pytorch/torch/csrc/autograd/autograd_not_implemented_fallback.cpp

#include <torch/csrc/autograd/autograd_not_implemented_fallback.h>

#include <c10/util/irange.h>

#include <ATen/core/TorchDispatchUtils.h>
#include <ATen/core/dispatch/Dispatcher.h>
#include <ATen/core/ivalue.h>

#include <c10/core/impl/TorchDispatchModeTLS.h>
#include <torch/csrc/autograd/VariableTypeUtils.h>
#include <torch/csrc/autograd/autograd.h>
#include <torch/csrc/autograd/function.h>
#include <torch/csrc/autograd/functions/basic_ops.h>
#include <torch/csrc/autograd/functions/utils.h>

#include <optional>
#include <utility>
#include <vector>

namespace torch {
namespace autograd {

namespace {

template <typename F>
void _foreach_tensor(
    F fn,
    torch::jit::Stack* stack,
    size_t stack_start,
    size_t size) {
  // Enumerate over tensors in a stack, including ones in TensorLists
  int idx_tensor = 0;
  for (const auto idx_arg : c10::irange(size)) {
    auto& ivalue = (*stack)[stack_start + idx_arg];
    if (ivalue.isTensor()) { // true for optional tensor that has value
      const auto& tensor = ivalue.toTensor();
      fn(idx_tensor, idx_arg, tensor);
      idx_tensor++;
    } else if (ivalue.isTensorList()) {
      for (const auto& iv : ivalue.toListRef()) {
        const auto& tensor = iv.toTensor();
        fn(idx_tensor, idx_arg, tensor);
        idx_tensor++;
      }
    }
  }
}

AutogradFallbackMode kAutogradFallbackMode = AutogradFallbackMode::Warn;

} // namespace

void setAutogradFallbackMode(AutogradFallbackMode mode) {
  TORCH_CHECK(mode != AutogradFallbackMode::Error, "NYI: mode='error'");
  kAutogradFallbackMode = mode;
}

AutogradFallbackMode getAutogradFallbackMode() {
  return kAutogradFallbackMode;
}

static void warnAutogradNotImplemented(const std::string& op_name) {
  TORCH_WARN(
      op_name,
      ": an autograd kernel was not registered to the Autograd key(s) ",
      "but we are trying to backprop through it. This may lead to silently incorrect behavior. ",
      "This behavior is deprecated and will be removed in a future version of PyTorch. ",
      "If your operator is differentiable, please ensure you have registered an "
      "autograd kernel to the correct Autograd key (e.g. DispatchKey::Autograd, "
      "DispatchKey::CompositeImplicitAutograd). If your operator is not "
      "differentiable, or to squash this warning and use the previous behavior, "
      "please register torch::CppFunction::makeFallthrough() to DispatchKey::Autograd.");
}

struct WarnNotImplemented : public Node {
  WarnNotImplemented(
      std::string op_name,
      size_t num_outputs,
      edge_list&& next_edges)
      : Node(std::move(next_edges)),
        op_name(std::move(op_name)),
        num_outputs(num_outputs) {}

  WarnNotImplemented(std::string op_name, size_t num_outputs)
      : op_name(std::move(op_name)), num_outputs(num_outputs) {}

  variable_list apply(variable_list&& inputs) override;

  std::string op_name;
  size_t num_outputs;
};

auto WarnNotImplemented::apply(variable_list&& inputs) -> variable_list {
  warnAutogradNotImplemented(op_name);
  std::vector<at::Tensor> output(num_outputs);
  return output;
}

static void basicAutogradNotImplementedFallbackImpl(
    const c10::OperatorHandle& op,
    c10::DispatchKeySet dispatch_keys,
    torch::jit::Stack* stack) {
  const auto& schema = op.schema();
  const auto& op_name = schema.operator_name().name;
  const auto num_arguments = schema.arguments().size();
  const auto num_returns = schema.returns().size();
  const auto stack_start = stack->size() - num_arguments;

  if (getAutogradFallbackMode() == AutogradFallbackMode::Nothing) {
    op.redispatchBoxed(dispatch_keys & c10::after_autograd_keyset, stack);
    return;
  }
  TORCH_INTERNAL_ASSERT(
      getAutogradFallbackMode() == AutogradFallbackMode::Warn);

  bool any_input_requires_grad = false;
  _foreach_tensor(
      [&](size_t _, size_t idx_arg, const at::Tensor& t) {
        if (t.requires_grad()) {
          any_input_requires_grad = true;
        }
      },
      stack,
      stack_start,
      num_arguments);
  // Optimization: TLS access can be slow. So we only check if it necessary
  // by putting it after the requires_grad checks.
  any_input_requires_grad = any_input_requires_grad && GradMode::is_enabled();

  std::shared_ptr<WarnNotImplemented> grad_fn;
  if (any_input_requires_grad) {
    // NB: It is standard to collect edges from all tensors
    // (see generated/VariableTypeEverything.cpp for examples)
    std::vector<const at::Tensor*> all_tensors_on_stack;
    _foreach_tensor(
        [&](size_t _, size_t idx_arg, const at::Tensor& t) {
          all_tensors_on_stack.push_back(&t);
        },
        stack,
        stack_start,
        num_arguments);
    grad_fn = std::shared_ptr<WarnNotImplemented>(
        new WarnNotImplemented(op_name, all_tensors_on_stack.size()),
        deleteNode);
    grad_fn->set_next_edges(collect_next_edges(all_tensors_on_stack));
  }

  op.redispatchBoxed(dispatch_keys & c10::after_autograd_keyset, stack);

  if (any_input_requires_grad) {
    // NB: if the operator mutates any inputs in-place and does not return them
    // as outputs, we are unable to lazily raise a warning. This is OK because
    // we don't expect many existing operators to do this because of the amount
    // of technical expertise necessary (you would need to manually register an
    // autograd kernel without using autograd.Function)
    _foreach_tensor(
        [&](size_t _, size_t idx_ret, const at::Tensor& t) {
          if (!isDifferentiableType(t.scalar_type())) {
            return;
          }
          const bool is_mutable_output =
              schema.is_aliasing({c10::SchemaArgType::output, idx_ret}) &&
              schema.is_mutable({c10::SchemaArgType::output, idx_ret});

          // If the post-autograd implementation returns Tensors that require
          // grad, then we install a hook that will warn during the backwards.
          //
          // NB: If the operation is inplace and the inputs were views,
          // it is possible that the history was rebased and the hook will
          // not warn in all places where it should. That is, the following
          // won't warn:
          // >>> x = torch.randn(3, 3, requires_grad=True)
          // >>> z = x.clone()
          // >>> w = z[0]
          // >>> k = w[0]
          // >>> y = op(k)
          // >>> torch.autograd.grad(z.sum(), w)
          if (t.requires_grad()) {
            t.register_hook([op_name](const at::Tensor& grad) {
              warnAutogradNotImplemented(op_name);
            });
            // If history is rebased, then we will attempt to warn
            // on the view's base. This will catch most cases (because
            // users typically call .backward() and backprop through
            // the entire program).
            if (t.is_view() && is_mutable_output) {
              // NOLINTNEXTLINE(cppcoreguidelines-pro-type-const-cast)
              auto& base = const_cast<at::TensorBase&>(t._base());
              if (base.requires_grad()) {
                // Can only register_hook on tensors that require grad.
                base.register_hook([op_name](const at::TensorBase& grad) {
                  warnAutogradNotImplemented(op_name);
                });
              }
            }
            return;
          }

          // If the post-autograd implementation returns any Tensors that
          // don't require grad, then we install the WarnNotImplemented grad_fn.
          // This grad_fn warns in backward and returns undefined tensor
          // gradients.
          //
          // NOTE [autograd fallback and in-place operations]
          // If the schema says the output is mutable, and the output
          // is an input, and the input is a view Tensor, then...
          // we're not sure if set_history is OK to do, so we just skip
          // adding the grad_fn. Builtin operators do rebase_history here,
          // but custom operators may have multiple Tensor(a!) returns,
          // rebase_history assumes single Tensor(a!) return, and in general
          // custom ops don't have a good in-place story.
          if (!is_mutable_output) {
            // NOLINTNEXTLINE(cppcoreguidelines-pro-type-const-cast)
            set_history(const_cast<at::Tensor&>(t), grad_fn);
          }
        },
        stack,
        stack->size() - num_returns,
        num_returns);
  }
}

torch::CppFunction basicAutogradNotImplementedFallback() {
  return torch::CppFunction::makeFromBoxedFunction<
      &basicAutogradNotImplementedFallbackImpl>();
}

void VariableHooks::basic_autograd_not_implemented_fallback(
    const c10::OperatorHandle& op,
    c10::DispatchKeySet dispatch_keys,
    torch::jit::Stack* stack) const {
  basicAutogradNotImplementedFallbackImpl(op, dispatch_keys, stack);
}

static void autogradNotImplementedFallbackImpl(
    const c10::OperatorHandle& op,
    c10::DispatchKeySet dispatch_keys,
    torch::jit::Stack* stack) {
  // Mimics a subset of the logic of a VariableType NotImplemented kernel
  // See gen_variable_type.py
  const auto& schema = op.schema();
  const auto& op_name = schema.operator_name().name;
  const auto num_arguments = schema.arguments().size();
  const auto num_returns = schema.returns().size();
  const auto stack_start = stack->size() - num_arguments;
  const bool grad_mode = GradMode::is_enabled();
  std::vector<const at::Tensor*> tensors_requiring_grad_on_stack;

  // Keep track of which outputs are output of in-place modification
  // so we can rebase_history if necessary
  std::vector<bool> is_inplace_output(num_returns, false);
  bool any_is_inplace_output = false;
  std::vector<bool> is_aliased_output(num_returns, false);
  std::optional<size_t> aliased_output_idx;

  for (const auto i : c10::irange(num_returns)) {
    if (schema.is_aliasing({c10::SchemaArgType::output, i})) {
      if (schema.is_mutable({c10::SchemaArgType::output, i})) {
        is_inplace_output[i] = true;
        any_is_inplace_output = true;
      } else {
        TORCH_CHECK(
            !aliased_output_idx.has_value(),
            "Expected only a single output in the operator schema to have a non-write alias annotation (i.e., 'Tensor(a)'). "
            "Non-composite functions where multiple outputs are aliased with inputs aren't supported."
            "Please rewrite your function as a composite function.");
        aliased_output_idx = i;
      }
      is_aliased_output[i] = true;
    }
  }

  int64_t aliased_input_idx = -1;
  for (const auto i : c10::irange(num_arguments)) {
    if (schema.is_aliasing({c10::SchemaArgType::input, i}) &&
        !schema.is_mutable({c10::SchemaArgType::input, i})) {
      TORCH_CHECK(
          aliased_input_idx == -1,
          "Expected only a single input in the operator schema to have a non-write alias annotation (i.e., 'Tensor(a)'). "
          "Non-composite functions where multiple inputs are aliased with outputs aren't supported. "
          "Please rewrite your function as a composite function.");
      aliased_input_idx = static_cast<int64_t>(i);
    }
  }

  size_t num_tensor_inputs = 0; // Only used for DEBUG-only checks
  _foreach_tensor(
      [&](size_t _, size_t idx_arg, const at::Tensor& t) {
        if (grad_mode && t.requires_grad()) {
          tensors_requiring_grad_on_stack.push_back(&t);
        }
        num_tensor_inputs++;
        TORCH_CHECK_NOT_IMPLEMENTED(
            !isFwGradDefined(t),
            "Trying to use forward AD with ",
            op_name,
            " that does not support it.");
      },
      stack,
      stack_start,
      num_arguments);

  const bool any_requires_grad = !tensors_requiring_grad_on_stack.empty();

  _foreach_tensor(
      [&](size_t _, size_t i, const at::Tensor& t) {
        if (schema.is_mutable({c10::SchemaArgType::input, i})) {
          check_inplace(t, any_requires_grad);
        }
      },
      stack,
      stack_start,
      num_arguments);

  std::shared_ptr<NotImplemented> grad_fn;
  if (any_requires_grad) {
    grad_fn = std::shared_ptr<NotImplemented>(
        new NotImplemented(op_name), deleteNode);
    grad_fn->set_next_edges(
        collect_next_edges(tensors_requiring_grad_on_stack));
  }

#ifndef NDEBUG
  // See NOTE [ TensorImpl and Storage Pointer Sanity Checks ]
  auto stack_args_copy =
      std::vector<c10::IValue>(stack->begin() + stack_start, stack->end());
  std::vector<c10::intrusive_ptr<c10::TensorImpl>> impl_saved;
  impl_saved.reserve(num_tensor_inputs);
  std::vector<c10::optional<c10::Storage>> storage_saved;
  storage_saved.reserve(num_tensor_inputs);
  _foreach_tensor(
      [&](size_t idx, size_t _, const at::Tensor& t) {
        storage_saved.push_back(
            t.has_storage() ? c10::optional<c10::Storage>(t.storage())
                            : c10::nullopt);
        impl_saved.push_back(t.getIntrusivePtr());
      },
      &stack_args_copy,
      0,
      num_arguments);
#endif
  if (aliased_input_idx != -1 || any_is_inplace_output) {
    at::AutoDispatchBelowAutograd guard;
    op.redispatchBoxed(dispatch_keys & c10::after_autograd_keyset, stack);
  } else {
    // If neither in-place nor view
    at::AutoDispatchBelowADInplaceOrView guard;
    op.redispatchBoxed(
        dispatch_keys & c10::after_ADInplaceOrView_keyset, stack);
  }
#ifndef NDEBUG
  _foreach_tensor(
      [&](size_t idx_tensor, size_t _, const at::Tensor& t) {
        if (storage_saved.at(idx_tensor).has_value())
          TORCH_INTERNAL_ASSERT(
              storage_saved.at(idx_tensor).value().is_alias_of(t.storage()),
              op_name);
        if (impl_saved.at(idx_tensor))
          TORCH_INTERNAL_ASSERT(
              impl_saved.at(idx_tensor) == t.getIntrusivePtr(), op_name);
      },
      &stack_args_copy,
      0,
      num_arguments);
  _foreach_tensor(
      [&](size_t idx_tensor, size_t idx_ret, const at::Tensor& t) {
        if (at::impl::tensor_has_dispatch(t) ||
            at::impl::dispatch_mode_enabled())
          return;
        if (!is_inplace_output[idx_ret])
          TORCH_INTERNAL_ASSERT(
              t.use_count() <= 1, op_name); // Okay to return undefined tensor
        // note(crcrpar): `_foreach_norm` returns a list of scalar Tensors and
        // each Tensor shares a storage of a hidden, intermediate 1D Tensor
        // created inside the CUDA implementation. This is because the
        // reference implementation of nvidia/apex repo returns this 1D Tensor
        // where each element represents the norm of corresponding input Tensor,
        // here I want to return the same number of Tensors as the input
        // TensorList, see https://github.com/pytorch/pytorch/issues/93940
        if (!is_aliased_output[idx_ret] && t.has_storage() &&
            op_name != "aten::_foreach_norm")
          TORCH_INTERNAL_ASSERT(t.storage().use_count() == 1);
      },
      stack,
      stack->size() - num_returns,
      num_returns);
  // There should be only a single base-view pair, make sure their storage is
  // aliased.
  if (aliased_input_idx != -1 && aliased_output_idx.has_value()) {
    const c10::IValue& aliased_input_iv = stack_args_copy[aliased_input_idx];
    const c10::IValue& aliased_output_iv =
        (*stack)[stack->size() - num_returns + *aliased_output_idx];
    TORCH_INTERNAL_ASSERT(aliased_input_iv.isTensor(), op_name);
    TORCH_INTERNAL_ASSERT(
        aliased_output_iv.isTensor() || aliased_output_iv.isTensorList(),
        op_name);
    const at::Tensor& aliased_input = aliased_input_iv.toTensor();
    if (aliased_input.has_storage()) {
      if (aliased_output_iv.isTensor()) {
        const at::Tensor& aliased_output = aliased_input_iv.toTensor();
        TORCH_INTERNAL_ASSERT(
            aliased_input.storage().is_alias_of(aliased_output.storage()),
            op_name);
      } else {
        const auto aliased_output_vec = aliased_output_iv.toTensorVector();
        for (const auto& aliased_output : aliased_output_vec) {
          TORCH_INTERNAL_ASSERT(
              aliased_input.storage().is_alias_of(aliased_output.storage()),
              op_name);
        }
      }
    }
  }
#endif

  if (any_requires_grad) {
    _foreach_tensor(
        [&](size_t idx_tensor, size_t idx_ret, const at::Tensor& t) {
          if (isDifferentiableType(t.scalar_type())) {
            if (is_inplace_output[idx_ret]) {
              // NOLINTNEXTLINE(cppcoreguidelines-pro-type-const-cast)
              rebase_history(const_cast<at::Tensor&>(t), grad_fn);
            } else {
              // NOLINTNEXTLINE(cppcoreguidelines-pro-type-const-cast)
              set_history(const_cast<at::Tensor&>(t), grad_fn);
            }
          }
        },
        stack,
        stack->size() - num_returns,
        num_returns);
  }
}

torch::CppFunction autogradNotImplementedFallback() {
  return torch::CppFunction::makeFromBoxedFunction<
      &autogradNotImplementedFallbackImpl>();
}

static void autogradNotImplementedInplaceOrViewFallbackImpl(
    const c10::OperatorHandle& op,
    c10::DispatchKeySet dispatch_keys,
    torch::jit::Stack* stack) {
  // Mimics a subset of the logic from ADInplaceOrViewType kernel:
  // - see gen_inplace_or_view_type.py
  // - this should only be used with autogradNotImplementedFallback above
  // - For more information see
  // https://pytorch.org/tutorials/advanced/dispatcher
  //
  // NOTE [ Limitations of ADInplaceOrView boxed kernel ]
  //
  // This op should only be used with autogradNotImplementedFallback kernel
  // because there is some logic we need specifically to enforce that even
  // if we do in-place on view's created in this kernel, the proper "derivative
  // is not implemented" error is still raised.
  //
  // Just like the codegened kernel, we try to enforce some things:
  // - For views: we enforce that the view relationship is between the first
  // input
  //   and the first output (which may be either Tensor or vec of Tensors
  // - For inplace (TODO?): enforce that the same op cannot be both a view and
  // inplace
  //   that is not allowed in the gen_inplace_or_view logic
  const auto& schema = op.schema();
  const auto& op_name = schema.operator_name().name;
  const auto num_arguments = schema.arguments().size();
  const auto num_returns = schema.returns().size();
  const auto stack_start = stack->size() - num_arguments;

  at::Tensor aliased_input;

  int64_t aliased_output_idx = -1;
  for (const auto i : c10::irange(num_returns)) {
    if (schema.is_aliasing({c10::SchemaArgType::output, i}) &&
        !schema.is_mutable({c10::SchemaArgType::output, i})) {
      TORCH_CHECK(
          aliased_output_idx == -1,
          "Fallback ADInplaceOrView kernel expects only a single output in the operator schema to have a "
          "non-write alias annotation (i.e., 'Tensor(a)'). "
          "Non-composite functions where multiple outputs are aliased with inputs aren't supported."
          "Please rewrite your function as a composite function.");
      aliased_output_idx = static_cast<int64_t>(i);
    }
  }

  std::optional<size_t> aliased_input_idx;
  for (const auto i : c10::irange(num_arguments)) {
    if (schema.is_aliasing({c10::SchemaArgType::input, i}) &&
        !schema.is_mutable({c10::SchemaArgType::input, i})) {
      TORCH_CHECK(
          !aliased_input_idx.has_value(),
          "Fallback ADInplaceOrView kernel expects only a single input in the operator schema to have a "
          "non-write alias annotation (i.e., 'Tensor(a)'). "
          "Non-composite functions where multiple inputs are aliased with outputs aren't supported. "
          "Please rewrite your function as a composite function.");
      aliased_input_idx = i;
      const c10::IValue& aliased_input_iv =
          (*stack)[stack_start + i]; // get a reference to an ivalue on the
                                     // stack
      TORCH_CHECK(aliased_input_iv.isTensor());
      aliased_input =
          aliased_input_iv.toTensor(); // TODO: Can we avoid saving this tensor
                                       // and incurring the refcount bump?
    }
  }
  // See NOTE [ Limitations of ADInplaceOrView boxed kernel ] above
  TORCH_CHECK(
      (!aliased_input_idx.has_value() && aliased_output_idx == -1) ||
          (aliased_input_idx.has_value() && aliased_input_idx.value() == 0 &&
           aliased_output_idx == 0),
      "Fallback ADInplaceOrView kernel can only create view relationships between the first "
      "input and the first output (the output can be a vector of tensors). Please change the "
      "order of your operator's parameters so that this is the case.");
  const bool is_view = aliased_input_idx.has_value();

  {
    at::AutoDispatchBelowADInplaceOrView guard;
    op.redispatchBoxed(
        dispatch_keys & c10::after_ADInplaceOrView_keyset, stack);
  }

  for (const auto i : c10::irange(num_returns)) {
    if (schema.is_mutable({c10::SchemaArgType::output, i})) {
      increment_version((*stack)[stack->size() - num_returns + i].toTensor());
    }
  }

  if (is_view) {
    c10::IValue& aliased_output_iv =
        (*stack)[stack->size() - num_returns + aliased_output_idx];

    // See NOTE [ View + Inplace detection ] for more details about this logic
    // We always need this view_func because otherwise if we do in-place
    // on this view, we would implicitly use AsStridedBackward instead
    // of the NotImplemented node. For the cross-dtype/non-strided
    // cases, we would create something like this anyway
    auto error_msg =
        ("Mutating the view " + op_name +
         "which does not have a derivative implemented is forbidden.");
    auto erroring_view_func = std::make_unique<ErroringViewFunc>(error_msg);

    const auto erroring_rev_view_func = [op_name = op_name](const at::Tensor&) {
      TORCH_CHECK(
          false,
          "Accessing the reverse view for ",
          op_name,
          " which does not have a derivative implemented is forbidden.");
      return at::Tensor();
    };

    if (aliased_output_iv.isTensorList()) {
      auto aliased_output = aliased_output_iv.toTensorVector();
      for (auto& sub_output : aliased_output) {
        as_view(
            /* base=*/aliased_input,
            /* tensor=*/sub_output,
            /* is_bw_differentiable=*/true,
            /* is_fw_differentiable=*/true,
            /* view_func=*/std::move(erroring_view_func),
            /* rev_view_func=*/erroring_rev_view_func,
            /* creation_meta=*/
            InferenceMode::is_enabled()
                ? CreationMeta::INFERENCE_MODE
                : (at::GradMode::is_enabled() ? CreationMeta::MULTI_OUTPUT_NODE
                                              : CreationMeta::NO_GRAD_MODE));
      }
      auto result = std::move(aliased_output);
      stack->at(stack->size() - num_returns + aliased_output_idx) = result;
    } else {
      TORCH_CHECK(aliased_output_iv.isTensor());
      auto result = as_view(
          /* base=*/aliased_input,
          /* tensor=*/std::move(aliased_output_iv).toTensor(),
          /* is_bw_differentiable=*/true,
          /* is_fw_differentiable=*/true,
          /* view_func=*/std::move(erroring_view_func),
          /* rev_view_func=*/erroring_rev_view_func,
          /* creation_meta=*/
          InferenceMode::is_enabled()
              ? CreationMeta::INFERENCE_MODE
              : (at::GradMode::is_enabled() ? CreationMeta::DEFAULT
                                            : CreationMeta::NO_GRAD_MODE));
      stack->at(stack->size() - num_returns + aliased_output_idx) =
          std::move(result);
    }
  }
}

torch::CppFunction autogradNotImplementedInplaceOrViewFallback() {
  return torch::CppFunction::makeFromBoxedFunction<
      &autogradNotImplementedInplaceOrViewFallbackImpl>();
}

} // namespace autograd
} // namespace torch