[T153220354] Fix header inclusions in c10 (#1541) (#101846)

Summary:
This is a re-attempt to land the iwyu header changes, by taking the diff from [PR 100304](https://github.com/pytorch/pytorch/pull/100304), and adding the bare minimal changes to make the diff build corectly in the internal builds.

X-link: https://github.com/facebookresearch/pytorch3d/pull/1541

X-link: https://github.com/fairinternal/pytorch3d/pull/44

- Re-work D45769819 to fix header inclusions in c10

Test Plan:
```
buck2 build --no-remote-cache mode/dev-nosan //caffe2/c10/...

buck2 build --no-remote-cache mode/dev-nosan //deeplearning/fbgemm/fbgemm_gpu/...

buck2 build mode/dev-nosan //vision/fair/pytorch3d/pytorch3d:_C
```

Reviewed By: malfet

Differential Revision: D45920611

Pull Request resolved: https://github.com/pytorch/pytorch/pull/101846
Approved by: https://github.com/malfet, https://github.com/Skylion007

aten/src/ATen/CachedTensorUtils.cpp

@@ -1,7 +1,7 @@
 #include <ATen/ATen.h>
-
 #include <ATen/CachedTensorUtils.h>
 
+#include <c10/util/flat_hash_map.h>
 
 namespace at {
 namespace caching {

aten/src/ATen/autocast_mode.cpp

@@ -4,6 +4,7 @@
 #include <exception>
 #include <mutex>
 #include <ATen/CachedTensorUtils.h>
+#include <c10/util/flat_hash_map.h>
 
 namespace at {
 namespace autocast {

aten/src/ATen/core/TensorBase.h

@@ -12,6 +12,7 @@
 #include <c10/core/UndefinedTensorImpl.h>
 #include <c10/core/WrapDimMinimal.h>
 #include <c10/util/Exception.h>
+#include <c10/util/ExclusivelyOwned.h>
 #include <c10/util/ExclusivelyOwnedTensorTraits.h>
 #include <c10/util/MaybeOwned.h>
 #include <c10/util/Optional.h>

aten/src/ATen/core/ivalue.h

@@ -14,6 +14,8 @@
 #include <c10/util/MaybeOwned.h>
 #include <c10/util/intrusive_ptr.h>
 #include <typeindex>
+#include <unordered_map>
+#include <unordered_set>
 #include <utility>
 
 namespace torch {

aten/src/ATen/core/ivalue_inl.h

@@ -22,6 +22,7 @@
 #include <c10/core/UndefinedTensorImpl.h>
 #include <c10/core/impl/DeviceGuardImplInterface.h>
 #include <c10/util/FunctionRef.h>
+#include <c10/util/Logging.h>
 #include <c10/util/hash.h>
 #include <c10/util/intrusive_ptr.h>
 #include <c10/util/irange.h>

aten/src/ATen/core/tensor_type.cpp

@@ -1,5 +1,6 @@
 #include <ATen/core/Tensor.h>
 #include <ATen/core/jit_type.h>
+#include <c10/core/GradMode.h>
 
 #include <utility>
 

aten/src/ATen/core/type.cpp

@@ -7,6 +7,7 @@
 #include <ATen/core/grad_mode.h>
 #include <ATen/core/jit_type.h>
 #include <c10/macros/Macros.h>
+#include <c10/util/flat_hash_map.h>
 #include <c10/util/irange.h>
 #include <array>
 #include <iostream>

aten/src/ATen/cuda/CUDAContext.h

@@ -14,6 +14,7 @@
 #include <ATen/Context.h>
 #include <c10/cuda/CUDAStream.h>
 #include <c10/cuda/CUDAFunctions.h>
+#include <c10/util/Logging.h>
 #include <ATen/cuda/Exceptions.h>
 
 namespace at {

aten/src/ATen/cuda/CublasHandlePool.cpp

@@ -3,7 +3,11 @@
 
 #include <c10/cuda/CUDACachingAllocator.h>
 
+#include <map>
+#include <memory>
 #include <regex>
+#include <string>
+#include <tuple>
 
 namespace at { namespace cuda {
 

aten/src/ATen/native/BatchLinearAlgebra.cpp

@@ -1,5 +1,6 @@
 #define TORCH_ASSERT_ONLY_METHOD_OPERATORS
 #include <ATen/core/Tensor.h>
+#include <ATen/core/grad_mode.h>
 #include <ATen/Dispatch.h>
 #include <ATen/Parallel.h>
 #include <ATen/TensorMeta.h>

aten/src/ATen/native/Convolution.cpp

@@ -9,6 +9,7 @@
 #include <ATen/native/cpu/DepthwiseConvKernel.h>
 #include <ATen/native/utils/ParamUtils.h>
 #include <ATen/native/xnnpack/Engine.h>
+#include <c10/core/GradMode.h>
 #include <c10/util/accumulate.h>
 #include <c10/util/irange.h>
 #include <c10/macros/Macros.h>

aten/src/ATen/native/DispatchStub.h

@@ -1,7 +1,7 @@
 #pragma once
 
 #include <c10/core/DeviceType.h>
-#include <c10/macros/Export.h>
+#include <c10/macros/Macros.h>
 
 #include <atomic>
 #include <utility>

aten/src/ATen/native/Distance.cpp

@@ -1,5 +1,6 @@
 #define TORCH_ASSERT_ONLY_METHOD_OPERATORS
 #include <ATen/core/Tensor.h>
+#include <ATen/core/grad_mode.h>
 #include <ATen/ExpandUtils.h>
 #include <ATen/NamedTensorUtils.h>
 #include <ATen/TensorOperators.h>

aten/src/ATen/native/RNN.cpp

@@ -9,6 +9,8 @@
 #include <ATen/native/quantized/PackedParams.h>
 #include <ATen/native/quantized/cpu/fbgemm_utils.h>
 #include <ATen/native/quantized/cpu/QnnpackUtils.h>
+#include <c10/core/GradMode.h>
+#include <c10/macros/Macros.h>
 #include <c10/util/irange.h>
 #include <torch/custom_class.h>
 #include <torch/library.h>

aten/src/ATen/native/UnaryOps.h

@@ -2,6 +2,7 @@
 
 #include <ATen/native/DispatchStub.h>
 #include <ATen/Generator.h>
+#include <c10/core/Scalar.h>
 #include <stdexcept>
 
 namespace at {

aten/src/ATen/native/mkldnn/RNN.cpp

@@ -6,6 +6,8 @@
 
 #include <ATen/TensorUtils.h>
 #include <ATen/Dispatch.h>
+#include <c10/core/GradMode.h>
+#include <c10/macros/Macros.h>
 #include <c10/util/Exception.h>
 
 #ifndef AT_PER_OPERATOR_HEADERS

aten/src/ATen/native/nested/NestedTensorMatmul.cpp

@@ -10,6 +10,7 @@
 #include <ATen/TensorOperators.h>
 #include <ATen/TensorUtils.h>
 #include <ATen/core/Tensor.h>
+#include <ATen/core/grad_mode.h>
 #include <ATen/native/layer_norm.h>
 #include <ATen/native/nested/NestedTensorUtils.h>
 

aten/src/ATen/native/quantized/QTensor.cpp

@@ -1,7 +1,6 @@
 #include <ATen/ATen.h>
 #include <ATen/NativeFunctions.h>
 #include <ATen/native/TensorIterator.h>
-#include <ATen/native/cpu/Loops.h>
 #include <ATen/native/quantized/cpu/QuantUtils.h>
 #include <ATen/quantized/QTensorImpl.h>
 #include <ATen/quantized/Quantizer.h>

aten/src/ATen/native/quantized/TensorCompare.cpp

@@ -1,6 +1,5 @@
 #include <ATen/ATen.h>
 #include <ATen/CPUApplyUtils.h>
-#include <ATen/Dispatch.h>
 #include <ATen/ExpandUtils.h>
 #include <ATen/NativeFunctions.h>
 #include <ATen/native/ReduceOpsUtils.h>

aten/src/ATen/native/sparse/SparseCsrTensorMath.cpp

@@ -4,6 +4,7 @@
 #include <ATen/Parallel.h>
 #include <ATen/SparseCsrTensorUtils.h>
 #include <ATen/core/Tensor.h>
+#include <ATen/core/grad_mode.h>
 #include <ATen/mkl/Sparse.h>
 #include <ATen/native/BinaryOps.h>
 #include <ATen/native/CPUBlas.h>

aten/src/ATen/native/transformers/cuda/sdp_utils.cpp

@@ -3,6 +3,7 @@
 #include <ATen/TensorSubclassLikeUtils.h>
 #include <ATen/TensorUtils.h>
 #include <ATen/core/Tensor.h>
+#include <ATen/core/grad_mode.h>
 #include <ATen/cuda/CUDAContext.h>
 #include <ATen/detail/CUDAHooksInterface.h>
 #include <ATen/native/DispatchStub.h>

aten/src/ATen/templates/TensorBody.h

@@ -21,6 +21,7 @@
 #include <c10/core/UndefinedTensorImpl.h>
 #include <c10/core/WrapDimMinimal.h>
 #include <c10/util/Exception.h>
+#include <c10/util/ExclusivelyOwned.h>
 #include <c10/util/Deprecated.h>
 #include <c10/util/MaybeOwned.h>
 #include <c10/util/Optional.h>

c10/CMakeLists.txt

@@ -54,6 +54,26 @@ target_compile_options(c10 PRIVATE "-DC10_BUILD_MAIN_LIB")
 if(${COMPILER_SUPPORTS_HIDDEN_VISIBILITY})
   target_compile_options(c10 PRIVATE "-fvisibility=hidden")
 endif()
+
+option(C10_USE_IWYU "Use include-what-you-use to clean up header inclusion" OFF)
+if(C10_USE_IWYU)
+  find_program(iwyu NAMES include-what-you-use)
+  if(iwyu)
+    set(iwyu_cmd
+        "include-what-you-use"
+        "-Xiwyu"
+        "--transitive_includes_only"
+        "-Xiwyu"
+        "--no_fwd_decls"
+        "-Xiwyu"
+        "--prefix_header_includes=keep"
+        "-Xiwyu"
+        "--mapping_file=${CMAKE_CURRENT_LIST_DIR}/../tools/iwyu/all.imp"
+        )
+    set_property(TARGET c10 PROPERTY CXX_INCLUDE_WHAT_YOU_USE ${iwyu_cmd})
+  endif()
+endif()
+
 if(WERROR)
   target_compile_options_if_supported(c10 PRIVATE "-Werror=sign-compare")
   target_compile_options_if_supported(c10 PRIVATE "-Werror=shadow")

c10/core/DefaultDtype.h

@@ -1,7 +1,7 @@
 #pragma once
 
 #include <c10/core/ScalarType.h>
-#include <c10/macros/Macros.h>
+#include <c10/macros/Export.h>
 
 namespace caffe2 {
 class TypeMeta;

c10/core/Device.cpp

@@ -1,12 +1,10 @@
 #include <c10/core/Device.h>
-#include <c10/macros/Macros.h>
 #include <c10/util/Exception.h>
 
 #include <algorithm>
 #include <array>
 #include <cctype>
 #include <exception>
-#include <ostream>
 #include <string>
 #include <vector>
 

c10/core/Device.h

@@ -1,7 +1,7 @@
 #pragma once
 
 #include <c10/core/DeviceType.h>
-#include <c10/macros/Macros.h>
+#include <c10/macros/Export.h>
 #include <c10/util/Exception.h>
 
 #include <cstddef>

c10/core/DeviceType.h

@@ -5,7 +5,7 @@
 // ATen/core (which would require a lot more build system hacking.)
 // If you modify me, keep me synchronized with that file.
 
-#include <c10/macros/Macros.h>
+#include <c10/macros/Export.h>
 
 #include <functional>
 #include <ostream>

c10/core/DispatchKey.h

@@ -1,7 +1,7 @@
 #pragma once
 
 #include <c10/core/DeviceType.h>
-#include <c10/macros/Macros.h>
+#include <c10/macros/Export.h>
 #include <ostream>
 #include <string>
 

c10/core/DispatchKeySet.cpp

@@ -1,6 +1,5 @@
 #include <c10/core/DispatchKeySet.h>
 #include <c10/util/irange.h>
-#include <iostream>
 
 namespace c10 {
 

c10/core/DispatchKeySet.h

@@ -3,6 +3,7 @@
 #include <c10/util/Exception.h>
 #include <c10/util/Metaprogramming.h>
 #include <c10/util/llvmMathExtras.h>
+#include <array>
 #include <ostream>
 
 namespace c10 {

c10/core/GeneratorImpl.cpp

@@ -8,6 +8,7 @@
 #ifndef _WIN32
 #include <fcntl.h>
 #include <unistd.h>
+#else
 #include <chrono>
 #endif
 

c10/core/GradMode.h

@@ -1,7 +1,7 @@
 #pragma once
 
 #include <c10/core/AutogradState.h>
-#include <c10/macros/Macros.h>
+#include <c10/macros/Export.h>
 
 namespace c10 {
 

c10/core/InferenceMode.h

@@ -1,9 +1,8 @@
 #pragma once
 
 #include <c10/core/AutogradState.h>
-#include <c10/core/GradMode.h>
 #include <c10/core/impl/LocalDispatchKeySet.h>
-#include <c10/macros/Macros.h>
+#include <c10/macros/Export.h>
 
 namespace c10 {
 

c10/core/SafePyObject.h

@@ -1,7 +1,7 @@
 #pragma once
 
 #include <c10/core/impl/PyInterpreter.h>
-#include <c10/macros/Macros.h>
+#include <c10/macros/Export.h>
 #include <c10/util/python_stub.h>
 
 namespace c10 {

c10/core/Scalar.h

@@ -1,9 +1,7 @@
 #pragma once
 
-#include <assert.h>
 #include <stdint.h>
 #include <stdexcept>
-#include <string>
 #include <type_traits>
 #include <utility>
 

c10/core/ScalarType.h

@@ -1,6 +1,7 @@
 #pragma once
 
 #include <c10/util/BFloat16.h>
+#include <c10/util/Deprecated.h>
 #include <c10/util/Exception.h>
 #include <c10/util/Half.h>
 #include <c10/util/bits.h>

c10/core/StorageImpl.h

@@ -1,7 +1,6 @@
 #pragma once
 
 #include <c10/core/Allocator.h>
-#include <c10/core/ScalarType.h>
 #include <c10/core/SymInt.h>
 #include <c10/core/impl/PyObjectSlot.h>
 

c10/core/SymFloat.h

@@ -7,7 +7,6 @@
 #include <c10/util/intrusive_ptr.h>
 
 #include <limits>
-#include <memory>
 
 namespace c10 {
 

c10/core/SymInt.cpp

@@ -3,9 +3,7 @@
 #include <c10/core/SymInt.h>
 #include <c10/core/SymNodeImpl.h>
 #include <c10/util/intrusive_ptr.h>
-#include <array>
 #include <functional>
-#include <utility>
 
 namespace c10 {
 

c10/core/SymInt.h

@@ -7,6 +7,7 @@
 #include <c10/util/Optional.h>
 
 #include <numeric>
+#include <type_traits>
 
 namespace c10 {
 

c10/core/SymNodeImpl.h

@@ -1,10 +1,9 @@
 #pragma once
 
-#include <c10/macros/Macros.h>
+#include <c10/macros/Export.h>
 #include <c10/util/ArrayRef.h>
 #include <c10/util/Exception.h>
 #include <c10/util/intrusive_ptr.h>
-#include <memory>
 
 namespace c10 {
 

c10/core/TensorImpl.cpp

@@ -1,13 +1,14 @@
 #include <c10/core/TensorImpl.h>
 
-#include <c10/core/Backend.h>
+#include <c10/core/CopyBytes.h>
 #include <c10/core/InferenceMode.h>
 #include <c10/core/SymIntArrayRef.h>
-#include <c10/core/WrapDimMinimal.h>
 #include <c10/core/impl/LocalDispatchKeySet.h>
 #include <c10/core/impl/PyInterpreter.h>
 #include <c10/core/impl/TorchDispatchModeTLS.h>
+#include <c10/util/Logging.h>
 #include <c10/util/Optional.h>
+#include <c10/util/accumulate.h>
 #include <c10/util/irange.h>
 
 #include <utility>

c10/core/TensorImpl.h

@@ -1,34 +1,30 @@
 #pragma once
 
-#include <c10/core/Backend.h>
-#include <c10/core/CopyBytes.h>
 #include <c10/core/DispatchKeySet.h>
 #include <c10/core/InferenceMode.h>
 #include <c10/core/MemoryFormat.h>
+#include <c10/core/ScalarTypeToTypeMeta.h>
 #include <c10/core/Storage.h>
 #include <c10/core/SymBool.h>
 #include <c10/core/SymIntArrayRef.h>
-#include <c10/core/TensorOptions.h>
 #include <c10/core/WrapDimMinimal.h>
-#include <c10/core/impl/LocalDispatchKeySet.h>
 #include <c10/core/impl/PyObjectSlot.h>
 #include <c10/core/impl/SizesAndStrides.h>
+#include <c10/macros/Macros.h>
 #include <c10/util/DimVector.h>
 #include <c10/util/Exception.h>
 #include <c10/util/Flags.h>
-#include <c10/util/Logging.h>
 #include <c10/util/Optional.h>
 #include <c10/util/accumulate.h>
 #include <c10/util/intrusive_ptr.h>
 #include <c10/util/irange.h>
-#include <c10/util/python_stub.h>
 #include <c10/util/safe_numerics.h>
+#include <c10/util/typeid.h>
 
 #include <algorithm>
 #include <atomic>
 #include <limits>
 #include <memory>
-#include <numeric>
 #include <type_traits>
 #include <utility>
 
@@ -57,11 +53,6 @@ class Tensor;
 class TensorBase;
 } // namespace at
 
-namespace c10 {
-class Scalar;
-struct Storage;
-} // namespace c10
-
 namespace c10 {
 
 /**
@@ -151,8 +142,6 @@ struct C10_API PlacementDeleteContext {
   }
 };
 
-struct TensorImpl;
-
 struct C10_API AutogradMetaInterface {
   virtual void set_requires_grad(
       bool requires_grad,
@@ -418,20 +407,6 @@ struct C10_API VariableVersion {
 // C10_TensorImpl_Size_Check_Dummy_Class
 struct C10_API TensorImpl;
 
-// Forward declaration needed because TensorImpl needs to be friends with
-// C10_TensorImpl_Size_Check_Dummy_Class in order to check the size
-// of its private fields.
-template <
-    size_t cplusplus,
-    size_t clang_ver_major,
-    size_t gcc_ver,
-    size_t gcc_ver_minor,
-    size_t nvcc,
-    size_t cuda_version,
-    size_t cuda_version_major,
-    size_t ptr_size>
-class C10_TensorImpl_Size_Check_Dummy_Class;
-
 /**
  * NOTE: Some TensorImpl methods are small and not overridden in the
  * PyTorch codebase itself, but may theoretically need to be

c10/core/TensorOptions.cpp

@@ -2,7 +2,6 @@
 
 #include <c10/core/Device.h>
 #include <c10/core/Layout.h>
-#include <c10/core/ScalarType.h>
 #include <c10/util/Optional.h>
 
 #include <iostream>

c10/core/TensorOptions.h

@@ -3,17 +3,14 @@
 #include <c10/core/Backend.h>
 #include <c10/core/DefaultDtype.h>
 #include <c10/core/Device.h>
-#include <c10/core/DispatchKeySet.h>
 #include <c10/core/Layout.h>
 #include <c10/core/MemoryFormat.h>
 #include <c10/core/ScalarType.h>
 #include <c10/core/ScalarTypeToTypeMeta.h>
 
 #include <c10/macros/Macros.h>
-#include <c10/util/C++17.h>
 #include <c10/util/Optional.h>
 
-#include <cstddef>
 #include <iosfwd>
 #include <utility>
 

c10/core/WrapDimMinimal.h

@@ -1,7 +1,6 @@
 #pragma once
 
 #include <c10/core/SymInt.h>
-#include <c10/util/Exception.h>
 
 namespace c10 {
 

c10/core/impl/GPUTrace.cpp

@@ -1,5 +1,3 @@
-#include <mutex>
-
 #include <c10/core/impl/GPUTrace.h>
 #include <c10/util/CallOnce.h>
 

c10/core/impl/HermeticPyObjectTLS.h

@@ -1,6 +1,6 @@
 #pragma once
 
-#include <c10/macros/Macros.h>
+#include <c10/macros/Export.h>
 #include <atomic>
 
 namespace c10 {

c10/core/impl/PyInterpreter.h

@@ -4,7 +4,7 @@
 #include <c10/core/Layout.h>
 #include <c10/core/MemoryFormat.h>
 #include <c10/core/SymIntArrayRef.h>
-#include <c10/macros/Macros.h>
+#include <c10/macros/Export.h>
 #include <c10/util/ArrayRef.h>
 #include <c10/util/intrusive_ptr.h>
 #include <c10/util/python_stub.h>
@@ -17,7 +17,6 @@ namespace c10 {
 struct IValue;
 class OperatorHandle;
 struct TensorImpl;
-struct SafePyObject;
 } // namespace c10
 
 namespace torch {

c10/core/impl/PythonDispatcherTLS.cpp

@@ -1,5 +1,4 @@
-#include <c10/core/DispatchKeySet.h>
-#include <c10/core/SafePyObject.h>
+#include <c10/core/DispatchKey.h>
 #include <c10/core/impl/LocalDispatchKeySet.h>
 #include <c10/core/impl/PythonDispatcherTLS.h>
 

c10/core/impl/PythonDispatcherTLS.h

@@ -1,8 +1,7 @@
 #pragma once
 
-#include <c10/core/SafePyObject.h>
-#include <c10/macros/Macros.h>
-#include <c10/util/Optional.h>
+#include <c10/core/impl/PyInterpreter.h>
+#include <c10/macros/Export.h>
 
 namespace c10 {
 namespace impl {

c10/core/impl/TorchDispatchModeTLS.cpp

@@ -1,4 +1,4 @@
-#include <c10/core/DispatchKeySet.h>
+#include <c10/core/DispatchKey.h>
 #include <c10/core/SafePyObject.h>
 #include <c10/core/impl/LocalDispatchKeySet.h>
 #include <c10/core/impl/TorchDispatchModeTLS.h>

c10/core/impl/TorchDispatchModeTLS.h

@@ -1,7 +1,7 @@
 #pragma once
 
 #include <c10/core/SafePyObject.h>
-#include <c10/macros/Macros.h>
+#include <c10/macros/Export.h>
 
 namespace c10 {
 namespace impl {

c10/core/impl/alloc_cpu.h

@@ -1,6 +1,6 @@
 #pragma once
 
-#include <c10/macros/Macros.h>
+#include <c10/macros/Export.h>
 
 #include <cstddef>
 

c10/core/thread_pool.cpp

@@ -1,4 +1,5 @@
 #include <c10/core/thread_pool.h>
+#include <c10/util/Logging.h>
 
 namespace c10 {
 

c10/mobile/CPUProfilingAllocator.cpp

@@ -1,5 +1,3 @@
-#include <climits>
-
 #include <c10/core/impl/alloc_cpu.h>
 #include <c10/mobile/CPUProfilingAllocator.h>
 #include <c10/util/irange.h>

c10/mobile/CPUProfilingAllocator.h

@@ -1,13 +1,8 @@
 #pragma once
 
-#include <algorithm>
-#include <deque>
-#include <memory>
-#include <mutex>
-
-#include <c10/util/Exception.h>
-#include <c10/util/SmallVector.h>
 #include <c10/util/flat_hash_map.h>
+#include <memory>
+#include <vector>
 
 namespace c10 {
 

c10/util/Exception.cpp

@@ -1,4 +1,3 @@
-#include <c10/util/Backtrace.h>
 #include <c10/util/Exception.h>
 #include <c10/util/Logging.h>
 #include <c10/util/Type.h>

c10/util/Exception.h

@@ -2,14 +2,11 @@
 #define C10_UTIL_EXCEPTION_H_
 
 #include <c10/macros/Macros.h>
-#include <c10/util/Deprecated.h>
 #include <c10/util/StringUtil.h>
 #include <c10/util/variant.h>
 
 #include <cstddef>
 #include <exception>
-#include <ostream>
-#include <sstream>
 #include <string>
 #include <vector>
 

c10/util/LeftRight.h

@@ -3,9 +3,7 @@
 #include <c10/util/Synchronized.h>
 #include <array>
 #include <atomic>
-#include <functional>
 #include <mutex>
-#include <shared_mutex>
 #include <thread>
 
 namespace c10 {

c10/util/Metaprogramming.h

@@ -2,7 +2,6 @@
 
 #include <c10/util/Array.h>
 #include <c10/util/TypeList.h>
-#include <array>
 #include <functional>
 #include <type_traits>
 

c10/util/Type.h

@@ -3,7 +3,9 @@
 
 #include <cstddef>
 #include <string>
+#ifdef __GXX_RTTI
 #include <typeinfo>
+#endif // __GXX_RTTI
 
 #include <c10/macros/Macros.h>
 

c10/util/TypeTraits.h

@@ -1,7 +1,6 @@
 #pragma once
 
 #include <c10/util/C++17.h>
-#include <functional>
 
 namespace c10 {
 namespace guts {

c10/util/intrusive_ptr.h

@@ -1,13 +1,10 @@
 #pragma once
 
-#include <c10/util/C++17.h>
 #include <c10/util/Exception.h>
-#include <c10/util/ExclusivelyOwned.h>
 #include <c10/util/MaybeOwned.h>
 #include <atomic>
 #include <climits>
 #include <memory>
-#include <stdexcept>
 
 namespace pybind11 {
 template <typename, typename...>

c10/util/numa.cpp

@@ -1,3 +1,4 @@
+#include <c10/util/Exception.h>
 #include <c10/util/numa.h>
 
 C10_DEFINE_bool(caffe2_cpu_numa_enabled, false, "Use NUMA whenever possible.");

c10/util/numa.h

@@ -1,6 +1,8 @@
 #pragma once
 
-#include <c10/util/Logging.h>
+#include <c10/macros/Export.h>
+#include <c10/util/Flags.h>
+#include <stddef.h>
 
 C10_DECLARE_bool(caffe2_cpu_numa_enabled);
 

c10/util/typeid.h

@@ -1,30 +1,17 @@
 #pragma once
 
 #include <atomic>
-#include <cassert>
-#include <complex>
 #include <cstdlib>
 #include <memory>
 #include <mutex>
 #include <type_traits>
-#include <unordered_map>
-#include <unordered_set>
 #include <vector>
-#ifdef __GXX_RTTI
-#include <typeinfo>
-#endif
-
-#include <exception>
 
 #include <c10/macros/Macros.h>
-#include <c10/util/Backtrace.h>
-#include <c10/util/C++17.h>
 #include <c10/util/Exception.h>
 #include <c10/util/IdWrapper.h>
-#include <c10/util/Type.h>
 #include <c10/util/TypeIndex.h>
 #include <c10/util/TypeTraits.h>
-#include <c10/util/flat_hash_map.h>
 
 #include <c10/core/ScalarType.h>
 #include <c10/util/irange.h>

torch/csrc/Module.cpp

@@ -17,6 +17,7 @@
 #include <ATen/dlpack.h>
 #include <ATen/native/ConvUtils.h>
 #include <c10/core/DispatchKeySet.h>
+#include <c10/util/Backtrace.h>
 #include <c10/util/Logging.h>
 #include <c10/util/irange.h>
 #include <libshm.h>

torch/csrc/api/include/torch/detail/TensorDataContainer.h

@@ -3,6 +3,7 @@
 #include <ATen/Dispatch.h>
 #include <ATen/ScalarOps.h>
 #include <ATen/core/Tensor.h>
+#include <ATen/core/grad_mode.h>
 
 #include <c10/util/irange.h>
 

torch/csrc/api/src/optim/schedulers/reduce_on_plateau_scheduler.cpp

@@ -1,5 +1,7 @@
 #include <torch/optim/schedulers/reduce_on_plateau_scheduler.h>
 
+#include <iomanip>
+
 namespace torch {
 namespace optim {
 

torch/csrc/autograd/forward_grad.h

@@ -1,6 +1,7 @@
 #pragma once
 
 #include <ATen/core/Tensor.h>
+#include <unordered_set>
 
 namespace torch {
 namespace autograd {

torch/csrc/autograd/input_buffer.cpp

@@ -5,6 +5,7 @@
 #include <ATen/SparseCsrTensorUtils.h>
 #include <ATen/TensorOperators.h>
 #include <ATen/TensorSubclassLikeUtils.h>
+#include <ATen/core/grad_mode.h>
 #include <ATen/native/SparseTensorUtils.h>
 
 #include <c10/core/DeviceGuard.h>

torch/csrc/jit/mobile/type_parser.h

@@ -2,6 +2,7 @@
 
 #include <ATen/core/dynamic_type.h>
 #include <ATen/core/jit_type.h>
+#include <unordered_set>
 
 namespace c10 {