Add files using upload-large-folder tool

.gitattributes

@@ -145,3 +145,4 @@ tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/_inductor/_
 .venv/lib/python3.11/site-packages/torch/_inductor/codegen/__pycache__/cpp_wrapper_cpu.cpython-311.pyc filter=lfs diff=lfs merge=lfs -text
 .venv/lib/python3.11/site-packages/torch/_inductor/codegen/__pycache__/cpp.cpython-311.pyc filter=lfs diff=lfs merge=lfs -text
 .venv/lib/python3.11/site-packages/torch/_inductor/codegen/__pycache__/common.cpython-311.pyc filter=lfs diff=lfs merge=lfs -text
+.venv/lib/python3.11/site-packages/torch/_inductor/fx_passes/__pycache__/split_cat.cpython-311.pyc filter=lfs diff=lfs merge=lfs -text

.venv/lib/python3.11/site-packages/torch/_inductor/fx_passes/__pycache__/split_cat.cpython-311.pyc

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b6d007391b31b1b010874c0dfd08680792c26a7542040197359230b54bc6612
+size 114085

.venv/lib/python3.11/site-packages/torch/include/ATen/AccumulateType.h

@@ -0,0 +1,173 @@
+#pragma once
+#include <ATen/Config.h>
+#include <c10/core/DeviceType.h>
+#include <c10/core/ScalarType.h>
+#include <c10/util/BFloat16.h>
+#include <c10/util/Float8_e4m3fn.h>
+#include <c10/util/Float8_e4m3fnuz.h>
+#include <c10/util/Float8_e5m2.h>
+#include <c10/util/Float8_e5m2fnuz.h>
+#include <c10/util/Half.h>
+
+// Defines the accumulation type for a scalar type.
+// Example:
+//   using accscalar_t = acc_type<scalar_t, /*is_cuda*/true>;
+//
+// Accumulation types are an important concept in numeric computing
+// because you frequently want to perform intermediate computations
+// at a higher precision than the input and output precision, to avoid
+// compounding internal rounding errors.  Accumulation is the most
+// well-known intermediate computation (it is of great importance for
+// sum reduction and matrix multiply, for example), but in PyTorch
+// acc_type ends up getting used for all sorts of other intermediate
+// computations, so it perhaps would be more accurately (ahem) called an
+// "accurate" type.  acc_type is especially important for reduced
+// precision operations like float16 and bfloat16, where relatively
+// benign looking inputs can easily end up overflowing/underflowing.
+//
+// acc_type is parametrized by whether or not you are running on CUDA
+// or not, because on CUDA double precision operations are expensive
+// and so by default, we don't actually want to use double as an
+// acc_type on CUDA.  A lot of things are typed out below, but
+// basically, the table is generated by a few rules:
+//
+//  If bool:
+//      Use 'bool' as acc_type.
+//  If floating point:
+//      If CUDA, use 'float' as acc_type (unless scalar_t is double),
+//      otherwise (CPU) use 'double'
+//  If integral:
+//      Use 'int64_t' as acc_type
+//
+// You're not forced to use this template; if you happen to know
+// something specific about your use case, you can specify your own
+// desired behavior.  This template, however, will give you a reasonable
+// default that will work for all dtypes supported in PyTorch.
+
+#if defined(__CUDACC__)
+#include <cuda.h>
+#include <cuda_fp16.h>
+#elif defined(__HIPCC__)
+#include <hip/hip_fp16.h>
+#include <hip/hip_runtime.h>
+#endif
+
+namespace at {
+
+template <typename T, c10::DeviceType D>
+struct AccumulateTypeDevice {};
+
+template <typename T, bool>
+struct AccumulateType {};
+
+template <typename T>
+struct AccumulateType<T, false> {
+  using type = typename AccumulateTypeDevice<T, c10::DeviceType::CPU>::type;
+};
+
+template <typename T>
+struct AccumulateType<T, true> {
+  using type = typename AccumulateTypeDevice<T, c10::DeviceType::CUDA>::type;
+};
+
+template <typename T, c10::DeviceType device>
+using acc_type_device = typename AccumulateTypeDevice<T, device>::type;
+
+template <typename T, bool is_cuda>
+using acc_type = typename AccumulateType<T, is_cuda>::type;
+
+#define ACC_TYPE(t, acc_t, device_type)         \
+  template <>                                   \
+  struct AccumulateTypeDevice<t, device_type> { \
+    using type = acc_t;                         \
+  };
+#define MPS_ACC_TYPE(t, acc_t) ACC_TYPE(t, acc_t, c10::DeviceType::MPS)
+#define XPU_ACC_TYPE(t, acc_t) ACC_TYPE(t, acc_t, c10::DeviceType::XPU)
+#define CUDA_ACC_TYPE(t, acc_t) ACC_TYPE(t, acc_t, c10::DeviceType::CUDA)
+#define CPU_ACC_TYPE(t, acc_t) ACC_TYPE(t, acc_t, c10::DeviceType::CPU)
+
+MPS_ACC_TYPE(BFloat16, float);
+MPS_ACC_TYPE(Half, float);
+MPS_ACC_TYPE(Float8_e5m2, float);
+MPS_ACC_TYPE(Float8_e4m3fn, float);
+MPS_ACC_TYPE(Float8_e5m2fnuz, float);
+MPS_ACC_TYPE(Float8_e4m3fnuz, float);
+MPS_ACC_TYPE(float, float);
+MPS_ACC_TYPE(double, float);
+MPS_ACC_TYPE(int8_t, int64_t);
+MPS_ACC_TYPE(uint8_t, int64_t);
+MPS_ACC_TYPE(char, int64_t);
+MPS_ACC_TYPE(int16_t, int64_t);
+MPS_ACC_TYPE(int32_t, int64_t);
+MPS_ACC_TYPE(int64_t, int64_t);
+MPS_ACC_TYPE(bool, bool);
+MPS_ACC_TYPE(c10::complex<Half>, c10::complex<float>);
+MPS_ACC_TYPE(c10::complex<float>, c10::complex<float>);
+MPS_ACC_TYPE(c10::complex<double>, c10::complex<float>);
+
+XPU_ACC_TYPE(BFloat16, float);
+XPU_ACC_TYPE(Half, float);
+XPU_ACC_TYPE(Float8_e5m2, float);
+XPU_ACC_TYPE(Float8_e4m3fn, float);
+XPU_ACC_TYPE(Float8_e5m2fnuz, float);
+XPU_ACC_TYPE(Float8_e4m3fnuz, float);
+XPU_ACC_TYPE(float, float);
+XPU_ACC_TYPE(double, double);
+XPU_ACC_TYPE(int8_t, int64_t);
+XPU_ACC_TYPE(uint8_t, int64_t);
+XPU_ACC_TYPE(char, int64_t);
+XPU_ACC_TYPE(int16_t, int64_t);
+XPU_ACC_TYPE(int32_t, int64_t);
+XPU_ACC_TYPE(int64_t, int64_t);
+XPU_ACC_TYPE(bool, bool);
+XPU_ACC_TYPE(c10::complex<Half>, c10::complex<float>);
+XPU_ACC_TYPE(c10::complex<float>, c10::complex<float>);
+XPU_ACC_TYPE(c10::complex<double>, c10::complex<double>);
+
+#if defined(__CUDACC__) || defined(__HIPCC__)
+CUDA_ACC_TYPE(half, float);
+#endif
+CUDA_ACC_TYPE(BFloat16, float);
+CUDA_ACC_TYPE(Half, float);
+CUDA_ACC_TYPE(Float8_e5m2, float);
+CUDA_ACC_TYPE(Float8_e4m3fn, float);
+CUDA_ACC_TYPE(Float8_e5m2fnuz, float);
+CUDA_ACC_TYPE(Float8_e4m3fnuz, float);
+CUDA_ACC_TYPE(float, float);
+CUDA_ACC_TYPE(double, double);
+CUDA_ACC_TYPE(int8_t, int64_t);
+CUDA_ACC_TYPE(uint8_t, int64_t);
+CUDA_ACC_TYPE(char, int64_t);
+CUDA_ACC_TYPE(int16_t, int64_t);
+CUDA_ACC_TYPE(int32_t, int64_t);
+CUDA_ACC_TYPE(int64_t, int64_t);
+CUDA_ACC_TYPE(bool, bool);
+CUDA_ACC_TYPE(c10::complex<Half>, c10::complex<float>);
+CUDA_ACC_TYPE(c10::complex<float>, c10::complex<float>);
+CUDA_ACC_TYPE(c10::complex<double>, c10::complex<double>);
+
+CPU_ACC_TYPE(BFloat16, float);
+CPU_ACC_TYPE(Half, float);
+CPU_ACC_TYPE(Float8_e5m2, float);
+CPU_ACC_TYPE(Float8_e4m3fn, float);
+CPU_ACC_TYPE(Float8_e5m2fnuz, float);
+CPU_ACC_TYPE(Float8_e4m3fnuz, float);
+CPU_ACC_TYPE(float, double);
+CPU_ACC_TYPE(double, double);
+CPU_ACC_TYPE(int8_t, int64_t);
+CPU_ACC_TYPE(uint8_t, int64_t);
+CPU_ACC_TYPE(char, int64_t);
+CPU_ACC_TYPE(int16_t, int64_t);
+CPU_ACC_TYPE(int32_t, int64_t);
+CPU_ACC_TYPE(int64_t, int64_t);
+CPU_ACC_TYPE(bool, bool);
+CPU_ACC_TYPE(c10::complex<Half>, c10::complex<float>);
+CPU_ACC_TYPE(c10::complex<float>, c10::complex<double>);
+CPU_ACC_TYPE(c10::complex<double>, c10::complex<double>);
+
+TORCH_API c10::ScalarType toAccumulateType(
+    c10::ScalarType type,
+    c10::DeviceType device);
+TORCH_API c10::ScalarType toAccumulateType(c10::ScalarType type, bool is_cuda);
+
+} // namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/Backend.h

@@ -0,0 +1,2 @@
+#pragma once
+#include <c10/core/Backend.h>

.venv/lib/python3.11/site-packages/torch/include/ATen/CPUApplyUtils.h

@@ -0,0 +1,343 @@
+#pragma once
+
+#include <ATen/CollapseDims.h>
+#include <ATen/Parallel.h>
+#include <ATen/TensorUtils.h>
+#include <c10/util/irange.h>
+#include <cstring>
+#include <limits>
+
+namespace at {
+
+/*
+ * The basic strategy for apply is as follows:
+ *
+ * 1. Starting with the outermost index, loop until we reach a dimension where
+ * the data is no longer contiguous, i.e. the stride at that dimension is not
+ * equal to the size of the tensor defined by the outer dimensions. Let's call
+ * this outer (contiguous) tensor A. Note that if the Tensor is contiguous, then
+ * A is equal to the entire Tensor. Let's call the inner tensor B.
+ *
+ * 2. We loop through the indices in B, starting at its outermost dimension. For
+ * example, if B is a 2x2 matrix, then we do:
+ *
+ * B[0][0]
+ * B[0][1]
+ * B[1][0]
+ * B[1][1]
+ *
+ * We set the offset into the underlying storage as (storageOffset + stride_B *
+ * index_B), i.e. basically we compute the offset into the storage as we would
+ * normally for a Tensor. But because we are guaranteed the subsequent data is
+ * contiguous in memory, we can simply loop for sizeof(A) iterations and perform
+ * the operation, without having to follow the order described by the strides of
+ * A.
+ *
+ * 3. As an optimization, we merge dimensions of A that are contiguous in
+ * memory. For example, if A is a 3x3x3x3 tensor narrowed from a 3x3x4x3 tensor,
+ * then the first two dimensions can be merged for the purposes of APPLY,
+ * reducing the number of nested loops.
+ */
+
+inline Tensor sort_strides(Tensor& tensor_) {
+  IntArrayRef strides = tensor_.strides();
+  std::vector<int64_t> indices;
+  indices.reserve(tensor_.ndimension());
+  for (const auto i : c10::irange(tensor_.ndimension())) {
+    indices.push_back(i);
+  }
+  std::sort(indices.begin(), indices.end(), [&strides](int64_t i1, int64_t i2) {
+    return strides[i1] > strides[i2];
+  });
+  Tensor tensor = tensor_.permute(indices);
+  return tensor;
+}
+
+template <typename T, int N>
+struct strided_tensor_iter_fixed {
+ public:
+  T* data_ = NULL;
+  int64_t dim_ = 0;
+
+  int64_t counter_[N] = {0};
+  int64_t sizes_[N] = {0};
+  int64_t strides_[N] = {0};
+
+  strided_tensor_iter_fixed(strided_tensor_iter_fixed const&) = delete;
+  void operator=(strided_tensor_iter_fixed const& x) = delete;
+  strided_tensor_iter_fixed(strided_tensor_iter_fixed&&) = default;
+  strided_tensor_iter_fixed(
+      Tensor& tensor,
+      C10_UNUSED bool sort_strides = false)
+      : data_(tensor.data_ptr<T>()) {
+    std::memset(counter_, 0, sizeof(int64_t) * N);
+    if (tensor.dim() > 0) {
+      std::memcpy(
+          sizes_, tensor.sizes().data(), tensor.dim() * sizeof(int64_t));
+      std::memcpy(
+          strides_, tensor.strides().data(), tensor.dim() * sizeof(int64_t));
+    }
+    dim_ = std::get<1>(collapse_dims(sizes_, strides_, tensor.ndimension()));
+  }
+};
+
+template <typename T>
+struct strided_tensor_iter {
+ private:
+ public:
+  T* data_ = NULL;
+  int64_t dim_;
+
+  std::vector<int64_t> counter_;
+  std::vector<int64_t> sizes_;
+  std::vector<int64_t> strides_;
+
+  strided_tensor_iter(strided_tensor_iter const&) = delete;
+  void operator=(strided_tensor_iter const& x) = delete;
+  strided_tensor_iter(strided_tensor_iter&&) = default;
+  strided_tensor_iter(Tensor& tensor)
+      : data_(tensor.data_ptr<T>()),
+        dim_(tensor.ndimension()),
+        counter_(dim_, 0),
+        sizes_(tensor.sizes().vec()),
+        strides_(tensor.strides().vec()) {
+    dim_ = std::get<1>(collapse_dims(sizes_.data(), strides_.data(), dim_));
+  }
+};
+
+inline bool _all_equal_numel(at::ArrayRef<Tensor> tensors) {
+  if (tensors.empty())
+    return true;
+  int64_t all_numel = tensors[0].numel();
+  for (const auto i : c10::irange(1, tensors.size())) {
+    if (tensors[i].numel() != all_numel)
+      return false;
+  }
+  return true;
+}
+
+inline std::string _all_equal_numel_error(at::ArrayRef<Tensor> tensors) {
+  std::ostringstream oss;
+  oss << "inconsistent tensor size, expected ";
+  for (size_t i = 0; i < tensors.size() - 1; i++) {
+    oss << tensors[i].sizes() << ", ";
+  }
+  oss << "and " << tensors[tensors.size() - 1].sizes()
+      << " to have the same number of elements, but got ";
+  for (size_t i = 0; i < tensors.size() - 1; i++) {
+    oss << tensors[i].numel() << ", ";
+  }
+  oss << "and " << tensors[tensors.size() - 1].numel()
+      << " elements respectively";
+  return oss.str();
+}
+
+inline bool _apply_preamble(ArrayRef<Tensor> tensors) {
+  checkDeviceType("CPU_tensor_apply", tensors, kCPU);
+  checkLayout("CPU_tensor_apply", tensors, kStrided);
+  if (!_all_equal_numel(tensors))
+    AT_ERROR(_all_equal_numel_error(tensors));
+  // An empty tensor has no elements
+  for (auto& t : tensors)
+    if (t.numel() == 0)
+      return false;
+  return true;
+}
+
+inline int64_t _max_dim_tensors(ArrayRef<Tensor> tensors) {
+  int64_t dim = 0;
+  for (auto& t : tensors)
+    dim = std::max(dim, t.ndimension());
+  return dim;
+}
+
+inline void iterate(int64_t /*size*/){};
+
+template <typename Arg, typename... Args>
+inline void iterate(int64_t size, Arg& iter, Args&... iter_tail) {
+  iter.counter_[iter.dim_ - 1] += size;
+  iter.data_ = iter.data_ + size * iter.strides_[iter.dim_ - 1];
+  iterate(size, iter_tail...);
+}
+
+inline bool iterate_continue() {
+  return true;
+};
+
+template <typename Arg, typename... Args>
+inline bool iterate_continue(Arg& iter, Args&... iter_tail) {
+  return iter.counter_[iter.dim_ - 1] < iter.sizes_[iter.dim_ - 1] &&
+      iterate_continue(iter_tail...);
+}
+
+inline int64_t max_iterate_size() {
+  return std::numeric_limits<int64_t>::max();
+};
+
+template <typename Arg, typename... Args>
+inline int64_t max_iterate_size(Arg& iter, Args&... iter_tail) {
+  return std::min(
+      (iter.sizes_[iter.dim_ - 1] - iter.counter_[iter.dim_ - 1]),
+      max_iterate_size(iter_tail...));
+}
+
+inline void iterate_overflow(){};
+
+template <typename Arg, typename... Args>
+inline void iterate_overflow(Arg& iter, Args&... iter_tail) {
+  if (iter.counter_[iter.dim_ - 1] == iter.sizes_[iter.dim_ - 1]) {
+    for (int64_t i = iter.dim_ - 1; i > 0; i--) {
+      if (iter.counter_[i] == iter.sizes_[i]) {
+        iter.counter_[i] = 0;
+        iter.counter_[i - 1]++;
+        iter.data_ = iter.data_ - (iter.sizes_[i] * iter.strides_[i]) +
+            iter.strides_[i - 1];
+      }
+    }
+  }
+  iterate_overflow(iter_tail...);
+}
+
+inline void forward(int64_t /*offset*/){};
+
+template <typename Arg, typename... Args>
+inline void forward(int64_t offset, Arg& iter, Args&... iter_tail) {
+  int64_t multi = offset;
+  for (int64_t i = iter.dim_ - 1; i >= 0; i--) {
+    int64_t inc = multi % iter.sizes_[i];
+    multi = multi / iter.sizes_[i];
+    iter.data_ = iter.data_ + inc * iter.strides_[i];
+    iter.counter_[i] += inc;
+  }
+  forward(offset, iter_tail...);
+}
+
+inline int64_t max_dim() {
+  return 0;
+}
+
+template <typename Arg, typename... Args>
+inline int64_t max_dim(Arg& iter, Args&... iter_tail) {
+  return std::max(iter.dim_, max_dim(iter_tail...));
+}
+
+inline void apply_op(){};
+
+template <typename Op, typename... Args>
+inline void apply_op(
+    int64_t numel,
+    int64_t offset,
+    const Op& op,
+    Args... iters) {
+  // For 0-dim tensors
+  if (numel == 1 && max_dim(iters...) == 0) {
+    op(*iters.data_...);
+    return;
+  }
+  if (offset > 0)
+    forward(offset, iters...);
+  // Splitting this into chunks helps the compiler create faster assembly
+  for (int64_t i = 0; i < numel;) {
+    for (; iterate_continue(iters...) && i < numel;) {
+      op(*iters.data_...);
+      iterate(1, iters...);
+      i++;
+    }
+    iterate_overflow(iters...);
+  }
+}
+
+/*
+  Apply a pointwise operator to sequence of tensors
+
+  The calling convention for op is a function/functor that takes the same
+  number of pointers of type scalar as the number of given tensors. For example,
+  to compute a = b * c, op would be of the form:
+  [](scalar* a_val, const scalar* b_val, const scalar* c_val) { a_val[0] =
+  b_val[0] * c_val[0]; };
+*/
+
+template <typename scalar1, typename scalar2, typename Op>
+inline void CPU_tensor_apply2(Tensor tensor1, Tensor tensor2, const Op op) {
+  if (!_apply_preamble({tensor1, tensor2}))
+    return;
+  if (_max_dim_tensors({tensor1, tensor2}) <= 8) {
+    apply_op(
+        tensor1.numel(),
+        0,
+        op,
+        strided_tensor_iter_fixed<scalar1, 8>(tensor1),
+        strided_tensor_iter_fixed<scalar2, 8>(tensor2));
+  } else {
+    apply_op(
+        tensor1.numel(),
+        0,
+        op,
+        strided_tensor_iter<scalar1>(tensor1),
+        strided_tensor_iter<scalar2>(tensor2));
+  }
+}
+
+template <typename scalar1, typename scalar2, typename scalar3, typename Op>
+inline void CPU_tensor_apply3(
+    Tensor tensor1,
+    Tensor tensor2,
+    Tensor tensor3,
+    const Op op) {
+  if (!_apply_preamble({tensor1, tensor2, tensor3}))
+    return;
+  if (_max_dim_tensors({tensor1, tensor2, tensor3}) <= 8) {
+    apply_op(
+        tensor1.numel(),
+        0,
+        op,
+        strided_tensor_iter_fixed<scalar1, 8>(tensor1),
+        strided_tensor_iter_fixed<scalar2, 8>(tensor2),
+        strided_tensor_iter_fixed<scalar3, 8>(tensor3));
+  } else {
+    apply_op(
+        tensor1.numel(),
+        0,
+        op,
+        strided_tensor_iter<scalar1>(tensor1),
+        strided_tensor_iter<scalar2>(tensor2),
+        strided_tensor_iter<scalar3>(tensor3));
+  }
+}
+
+template <
+    typename scalar1,
+    typename scalar2,
+    typename scalar3,
+    typename scalar4,
+    typename Op>
+inline void CPU_tensor_apply4(
+    Tensor tensor1,
+    Tensor tensor2,
+    Tensor tensor3,
+    Tensor tensor4,
+    const Op op) {
+  if (!_apply_preamble({tensor1, tensor2, tensor3, tensor4}))
+    return;
+  if (_max_dim_tensors({tensor1, tensor2, tensor3, tensor4}) <= 8) {
+    apply_op(
+        tensor1.numel(),
+        0,
+        op,
+        strided_tensor_iter_fixed<scalar1, 8>(tensor1),
+        strided_tensor_iter_fixed<scalar2, 8>(tensor2),
+        strided_tensor_iter_fixed<scalar3, 8>(tensor3),
+        strided_tensor_iter_fixed<scalar4, 8>(tensor4));
+  } else {
+    apply_op(
+        tensor1.numel(),
+        0,
+        op,
+        strided_tensor_iter<scalar1>(tensor1),
+        strided_tensor_iter<scalar2>(tensor2),
+        strided_tensor_iter<scalar3>(tensor3),
+        strided_tensor_iter<scalar4>(tensor4));
+  }
+}
+
+} // namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/CPUFixedAllocator.h

@@ -0,0 +1,33 @@
+#pragma once
+
+#include <c10/core/Allocator.h>
+#include <c10/util/Exception.h>
+
+// This file creates a fake allocator that just throws exceptions if
+// it is actually used.
+
+// state passed to the allocator is the std::function<void(void*)> called
+// when the blob is release by ATen
+
+namespace at {
+
+static cpu_fixed_malloc(void*, ptrdiff_t) {
+  AT_ERROR("attempting to resize a tensor view of an external blob");
+}
+
+static cpu_fixed_realloc(void*, void*, ptrdiff_t) {
+  AT_ERROR("attempting to resize a tensor view of an external blob");
+}
+
+static cpu_fixed_free(void* state, void* allocation) {
+  auto on_release = static_cast<std::function<void(void*)>*>(state);
+  (*on_release)(allocation);
+  delete on_release;
+}
+
+static Allocator CPU_fixed_allocator = {
+    cpu_fixed_malloc,
+    cpu_fixed_realloc,
+    cpu_fixed_free};
+
+} // namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/CPUFunctions.h

@@ -0,0 +1,29 @@
+#include <ATen/core/TensorBody.h>
+
+// TODO Undo all logic introduced for Note [Avoiding Include Cycles In Static Dispatch]
+// Code introduced to avoid cyclic dependency in static dispatch is no longer
+// needed as static dispatch logic is moved from TensorBody.h, which caused cycles in the first place,
+// to Operators.cpp for supporting multiple backends with multiple kernels.
+//
+// Note [Avoiding Include Cycles In Static Dispatch]
+// In order to avoid #include cycles in the static dispatch build, we've carefully split out
+// the static function definition files into {DispatchKey}Functions.h and {DispatchKey}Functions_inl.h.
+//
+// Without this split, the include cycle looks like TensorBody.h -> CPUFunctions.h -> TensorBody.h.
+// - TensorBody.h #includes CPUFunctions.h in the static dispatch build, because the tensor methods
+//   all need to call into the fastpath C++ API defined in CPUFunctions.h. The methods are also all
+//   directly inlined into TensorBody.h.
+// - CPUFunctions.h #includes TensorBody.h because it contains function declarations for the entire C++ API,
+//   which include functions that have defaultable std::optional<Tensor> arguments.
+//   That requires knowing the full Tensor class definition.
+//
+// We break the cycle by doing the following:
+// - Split out CPUFunction.h into two files: CPUFunctions.h and CPUFunctions_inl.h
+// - CPUFunction.h is a dummy file that just includes the Tensor class and includes CPUFunctions_inl.,
+// - CPUFunctions_inl.h includes everything else
+// - (only in the static dispatch build) TensorBody.h makes sure to finish defining the Tensor class,
+//   and then it includes CPUFunctions_inl.h.
+// - All other files that want the cpu fastpath functions can include CPUFunctions.h directly.
+// - This also means that static dispatch build, CPUFunctions.h only needs to
+//   #include TensorBody.h, and it will automatically bring in CPUFunctions_inl.h.
+#include <ATen/CPUFunctions_inl.h>

.venv/lib/python3.11/site-packages/torch/include/ATen/CPUGeneratorImpl.h

@@ -0,0 +1,49 @@
+#pragma once
+
+#include <ATen/core/Generator.h>
+#include <ATen/core/MT19937RNGEngine.h>
+#include <c10/core/GeneratorImpl.h>
+#include <optional>
+
+namespace at {
+
+struct TORCH_API CPUGeneratorImpl : public c10::GeneratorImpl {
+  // Constructors
+  CPUGeneratorImpl(uint64_t seed_in = default_rng_seed_val);
+  ~CPUGeneratorImpl() override = default;
+
+  // CPUGeneratorImpl methods
+  std::shared_ptr<CPUGeneratorImpl> clone() const;
+  void set_current_seed(uint64_t seed) override;
+  void set_offset(uint64_t offset) override;
+  uint64_t get_offset() const override;
+  uint64_t current_seed() const override;
+  uint64_t seed() override;
+  void set_state(const c10::TensorImpl& new_state) override;
+  c10::intrusive_ptr<c10::TensorImpl> get_state() const override;
+  static c10::DeviceType device_type();
+  uint32_t random();
+  uint64_t random64();
+  std::optional<float> next_float_normal_sample();
+  std::optional<double> next_double_normal_sample();
+  void set_next_float_normal_sample(std::optional<float> randn);
+  void set_next_double_normal_sample(std::optional<double> randn);
+  at::mt19937 engine();
+  void set_engine(at::mt19937 engine);
+
+ private:
+  CPUGeneratorImpl* clone_impl() const override;
+  at::mt19937 engine_;
+  std::optional<float> next_float_normal_sample_;
+  std::optional<double> next_double_normal_sample_;
+};
+
+namespace detail {
+
+TORCH_API const Generator& getDefaultCPUGenerator();
+TORCH_API Generator
+createCPUGenerator(uint64_t seed_val = default_rng_seed_val);
+
+} // namespace detail
+
+} // namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/CompositeExplicitAutogradFunctions_inl.h

@@ -0,0 +1,553 @@
+#pragma once
+// @generated by torchgen/gen.py from DispatchKeyFunctions_inl.h
+
+// NB: The implementing C++ file is RegisterDispatchKey.cpp
+
+// The only #includes we need are for custom classes that have defaults in the C++ API
+#include <c10/core/MemoryFormat.h>
+#include <c10/core/Scalar.h>
+#include <ATen/core/Reduction.h>
+
+#if defined(AT_PER_OPERATOR_HEADERS) && defined(TORCH_ASSERT_ONLY_METHOD_OPERATORS)
+#error This change adds a dependency on all pytorch operators, meaning the     \
+  file will need to be re-compiled every time an operator is changed or added. \
+  Consider including a specific operator from                                  \
+  <ATen/ops/{my_operator}_compositeexplicitautograd_dispatch.h>.                   \
+  See NOTE [TORCH_ASSERT_ONLY_METHOD_OPERATORS].
+#endif
+
+#include <ATen/ops/_adaptive_avg_pool2d_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_adaptive_avg_pool2d_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_adaptive_avg_pool3d_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_adaptive_avg_pool3d_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_add_relu_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_aminmax_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_amp_foreach_non_finite_check_and_unscale_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_amp_update_scale_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_assert_scalar_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_batch_norm_no_update_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_batch_norm_with_update_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_cdist_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_cdist_forward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_cholesky_solve_helper_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_chunk_cat_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_coalesce_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_coalesced_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_conj_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_conj_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_conj_physical_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_convolution_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_copy_from_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_copy_from_and_resize_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_ctc_loss_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_ctc_loss_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_cudnn_ctc_loss_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_cudnn_init_dropout_state_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_cudnn_rnn_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_cudnn_rnn_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_cudnn_rnn_flatten_weight_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_dirichlet_grad_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_efficientzerotensor_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_embedding_bag_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_embedding_bag_dense_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_embedding_bag_forward_only_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_embedding_bag_per_sample_weights_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_empty_affine_quantized_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_empty_per_channel_affine_quantized_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_euclidean_dist_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_fake_quantize_learnable_per_channel_affine_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_fake_quantize_learnable_per_tensor_affine_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_fake_quantize_per_tensor_affine_cachemask_tensor_qparams_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foobar_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_abs_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_acos_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_add_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_addcdiv_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_addcmul_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_asin_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_atan_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_ceil_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_clamp_max_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_clamp_min_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_cos_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_cosh_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_div_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_erf_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_erfc_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_exp_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_expm1_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_floor_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_frac_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_lerp_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_lgamma_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_log_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_log10_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_log1p_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_log2_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_max_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_maximum_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_minimum_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_mul_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_neg_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_norm_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_pow_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_reciprocal_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_round_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_sigmoid_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_sign_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_sin_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_sinh_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_sqrt_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_sub_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_tan_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_tanh_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_trunc_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_foreach_zero_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_functional_assert_scalar_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_functional_sym_constrain_range_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_functional_sym_constrain_range_for_size_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_fused_adagrad_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_fused_adam_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_fused_adamw_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_fused_dropout_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_fused_moving_avg_obs_fq_helper_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_fused_sgd_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_fw_primal_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_fw_primal_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_grid_sampler_2d_cpu_fallback_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_has_same_storage_numel_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_histogramdd_bin_edges_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_histogramdd_from_bin_cts_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_histogramdd_from_bin_tensors_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_index_put_impl_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_indices_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_is_all_true_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_is_any_true_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_lazy_clone_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_linalg_check_errors_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_lstm_mps_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_make_dual_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_make_dual_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_make_per_channel_quantized_tensor_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_make_per_tensor_quantized_tensor_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_masked_scale_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_masked_softmax_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_masked_softmax_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_mkldnn_reshape_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_mkldnn_transpose_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_mps_convolution_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_mps_convolution_transpose_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_native_batch_norm_legit_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_native_batch_norm_legit_no_training_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_native_multi_head_attention_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_neg_view_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_neg_view_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_nested_from_padded_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_nested_from_padded_and_nested_example_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_nested_get_values_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_nested_tensor_from_mask_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_nested_tensor_from_tensor_list_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_nested_tensor_size_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_nested_tensor_storage_offsets_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_nested_tensor_strides_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_nested_view_from_buffer_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_nested_view_from_jagged_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_new_zeros_with_same_feature_meta_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_nnpack_spatial_convolution_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_pack_padded_sequence_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_pdist_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_pdist_forward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_pin_memory_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_print_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_reshape_alias_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_reshape_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_resize_output_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_safe_softmax_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_sample_dirichlet_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_scaled_dot_product_fused_attention_overrideable_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_scaled_dot_product_fused_attention_overrideable_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_segment_reduce_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_slow_conv2d_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_addmm_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_broadcast_to_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_compressed_tensor_with_dims_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_coo_tensor_with_dims_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_coo_tensor_with_dims_and_tensors_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_csr_prod_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_csr_sum_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_log_softmax_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_log_softmax_backward_data_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_mask_projection_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_softmax_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_softmax_backward_data_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_sparse_matmul_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_sum_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_sum_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_spdiags_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_stack_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_standard_gamma_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_standard_gamma_grad_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_test_autograd_multiple_dispatch_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_test_autograd_multiple_dispatch_view_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_test_autograd_multiple_dispatch_view_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_test_functorch_fallback_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_test_optional_filled_intlist_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_test_optional_floatlist_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_test_optional_intlist_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_test_parallel_materialize_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_test_warn_in_autograd_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_thnn_fused_gru_cell_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_thnn_fused_gru_cell_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_thnn_fused_lstm_cell_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_thnn_fused_lstm_cell_backward_impl_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_to_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_to_dense_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_to_sparse_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_to_sparse_bsc_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_to_sparse_bsr_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_to_sparse_csc_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_to_sparse_csr_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_transform_bias_rescale_qkv_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_transformer_encoder_layer_fwd_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_trilinear_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_triton_multi_head_attention_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_triton_scaled_dot_attention_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_unique_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_unique2_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_unsafe_index_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_unsafe_index_put_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_unsafe_masked_index_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_unsafe_masked_index_put_accumulate_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_unsafe_view_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_values_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_weight_norm_interface_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/_weight_norm_interface_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/abs_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/add_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/addr_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/affine_grid_generator_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/alias_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/alias_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/all_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/allclose_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/any_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/arange_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/as_strided_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/as_strided_scatter_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/bartlett_window_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/batch_norm_backward_elemt_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/batch_norm_backward_reduce_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/batch_norm_gather_stats_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/batch_norm_gather_stats_with_counts_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/batch_norm_stats_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/batch_norm_update_stats_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/bernoulli_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/binary_cross_entropy_with_logits_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/bincount_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/binomial_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/bitwise_and_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/bitwise_left_shift_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/bitwise_or_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/bitwise_right_shift_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/bitwise_xor_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/blackman_window_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/block_diag_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/bucketize_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/cauchy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/ccol_indices_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/ccol_indices_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/celu_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/channel_shuffle_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/cholesky_solve_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/clone_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/col_indices_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/col_indices_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/complex_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/conj_physical_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/constant_pad_nd_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/conv_depthwise3d_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/conv_tbc_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/convolution_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/convolution_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/convolution_backward_overrideable_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/convolution_overrideable_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/copy_sparse_to_sparse_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/copysign_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/count_nonzero_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/crow_indices_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/crow_indices_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/cudnn_affine_grid_generator_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/cudnn_affine_grid_generator_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/cudnn_batch_norm_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/cudnn_batch_norm_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/cudnn_convolution_add_relu_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/cudnn_convolution_relu_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/cudnn_convolution_transpose_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/cudnn_grid_sampler_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/cudnn_grid_sampler_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/cummax_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/cummin_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/deg2rad_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/dense_dim_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/dequantize_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/detach_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/detach_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/diag_embed_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/diagonal_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/diagonal_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/diagonal_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/diagonal_scatter_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/dist_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/div_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/dot_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/embedding_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/embedding_dense_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/embedding_renorm_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/empty_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/empty_like_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/empty_permuted_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/empty_quantized_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/empty_strided_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/expand_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/expand_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/exponential_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/eye_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/fake_quantize_per_channel_affine_cachemask_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/fake_quantize_per_tensor_affine_cachemask_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/fft_fftfreq_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/fft_rfftfreq_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/fill_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/flip_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/floor_divide_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/fmod_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/frexp_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/from_file_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/full_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/full_like_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/geometric_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/glu_backward_jvp_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/glu_jvp_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/grid_sampler_2d_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/grid_sampler_2d_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/grid_sampler_3d_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/grid_sampler_3d_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/hamming_window_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/hann_window_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/hardswish_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/huber_loss_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/index_fill_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/index_put_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/indices_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/indices_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/int_repr_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/is_coalesced_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/is_pinned_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/is_same_size_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/isinf_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/isnan_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/kaiser_window_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/kthvalue_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/lift_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/lift_fresh_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/lift_fresh_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_lstsq_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_matrix_exp_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_pinv_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/linear_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/linear_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/linspace_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/log_normal_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/log_softmax_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/logcumsumexp_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/logical_and_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/logical_not_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/logical_or_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/logical_xor_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/logspace_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/logsumexp_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/lshift_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/lstm_mps_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/masked_fill_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/masked_scatter_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/masked_scatter_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/matmul_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/max_pool2d_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/mean_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/median_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/miopen_batch_norm_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/miopen_batch_norm_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/miopen_convolution_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/miopen_convolution_transpose_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/miopen_depthwise_convolution_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/miopen_rnn_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/miopen_rnn_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/mkldnn_adaptive_avg_pool2d_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/mkldnn_convolution_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/mkldnn_linear_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/mkldnn_linear_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/mkldnn_linear_backward_input_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/mkldnn_linear_backward_weights_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/mkldnn_max_pool2d_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/mkldnn_max_pool2d_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/mkldnn_max_pool3d_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/mkldnn_max_pool3d_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/mkldnn_reorder_conv2d_weight_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/mkldnn_reorder_conv3d_weight_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/mkldnn_rnn_layer_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/mkldnn_rnn_layer_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/mode_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/mps_convolution_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/mps_convolution_transpose_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/mul_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/mv_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/mvlgamma_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/nan_to_num_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/nanmedian_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/native_batch_norm_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/native_dropout_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/native_dropout_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/native_group_norm_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/native_group_norm_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/native_layer_norm_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/native_layer_norm_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/native_norm_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/new_empty_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/new_empty_strided_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/new_full_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/new_ones_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/new_zeros_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/norm_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/normal_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/ones_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/ones_like_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/permute_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/permute_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/pixel_shuffle_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/pixel_unshuffle_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/poisson_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/polar_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/polygamma_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/prod_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/put_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/q_per_channel_scales_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/q_per_channel_zero_points_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/quantize_per_channel_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/quantize_per_tensor_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/quantize_per_tensor_dynamic_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/quantized_batch_norm_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/quantized_max_pool1d_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/quantized_max_pool2d_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/quantized_max_pool3d_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/rad2deg_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/rand_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/rand_like_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/randint_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/randint_like_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/randn_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/randn_like_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/random_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/randperm_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/range_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/relu_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/remainder_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/repeat_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/repeat_interleave_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/resize_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/resize_as_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/resize_as_sparse_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/roll_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/rot90_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/row_indices_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/row_indices_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/rrelu_with_noise_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/rshift_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/rsub_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/scalar_tensor_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/segment_reduce_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/select_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/select_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/select_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/select_scatter_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/set_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/slice_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/slice_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/slice_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/slice_inverse_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/slice_scatter_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/slow_conv_dilated2d_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/slow_conv_dilated3d_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/smooth_l1_loss_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/soft_margin_loss_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/soft_margin_loss_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/softmax_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/sort_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/sparse_compressed_tensor_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/sparse_coo_tensor_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/sparse_dim_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/sparse_mask_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/sparse_resize_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/sparse_resize_and_clear_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/special_chebyshev_polynomial_t_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/special_chebyshev_polynomial_u_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/special_chebyshev_polynomial_v_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/special_chebyshev_polynomial_w_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/special_hermite_polynomial_h_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/special_hermite_polynomial_he_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/special_laguerre_polynomial_l_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/special_legendre_polynomial_p_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_t_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_u_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_v_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_w_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/special_xlog1py_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/special_zeta_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/split_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/split_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/split_with_sizes_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/split_with_sizes_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/squeeze_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/squeeze_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/stack_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/std_mean_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/sub_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/sum_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/sym_constrain_range_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/sym_constrain_range_for_size_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/t_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/t_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/to_mkldnn_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/to_padded_tensor_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/trace_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/transpose_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/transpose_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/tril_indices_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/triu_indices_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/unbind_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/unbind_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/unfold_backward_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/unfold_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/uniform_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/unique_consecutive_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/unique_dim_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/unique_dim_consecutive_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/unsafe_split_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/unsafe_split_with_sizes_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/unsqueeze_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/unsqueeze_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/values_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/values_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/var_mean_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/vdot_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/view_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/view_as_complex_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/view_as_real_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/view_copy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/xlogy_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/zero_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/zeros_compositeexplicitautograd_dispatch.h>
+#include <ATen/ops/zeros_like_compositeexplicitautograd_dispatch.h>
+
+
+

.venv/lib/python3.11/site-packages/torch/include/ATen/CompositeImplicitAutogradFunctions.h

@@ -0,0 +1,29 @@
+#include <ATen/core/TensorBody.h>
+
+// TODO Undo all logic introduced for Note [Avoiding Include Cycles In Static Dispatch]
+// Code introduced to avoid cyclic dependency in static dispatch is no longer
+// needed as static dispatch logic is moved from TensorBody.h, which caused cycles in the first place,
+// to Operators.cpp for supporting multiple backends with multiple kernels.
+//
+// Note [Avoiding Include Cycles In Static Dispatch]
+// In order to avoid #include cycles in the static dispatch build, we've carefully split out
+// the static function definition files into {DispatchKey}Functions.h and {DispatchKey}Functions_inl.h.
+//
+// Without this split, the include cycle looks like TensorBody.h -> CPUFunctions.h -> TensorBody.h.
+// - TensorBody.h #includes CPUFunctions.h in the static dispatch build, because the tensor methods
+//   all need to call into the fastpath C++ API defined in CPUFunctions.h. The methods are also all
+//   directly inlined into TensorBody.h.
+// - CPUFunctions.h #includes TensorBody.h because it contains function declarations for the entire C++ API,
+//   which include functions that have defaultable std::optional<Tensor> arguments.
+//   That requires knowing the full Tensor class definition.
+//
+// We break the cycle by doing the following:
+// - Split out CPUFunction.h into two files: CPUFunctions.h and CPUFunctions_inl.h
+// - CPUFunction.h is a dummy file that just includes the Tensor class and includes CPUFunctions_inl.,
+// - CPUFunctions_inl.h includes everything else
+// - (only in the static dispatch build) TensorBody.h makes sure to finish defining the Tensor class,
+//   and then it includes CPUFunctions_inl.h.
+// - All other files that want the cpu fastpath functions can include CPUFunctions.h directly.
+// - This also means that static dispatch build, CPUFunctions.h only needs to
+//   #include TensorBody.h, and it will automatically bring in CPUFunctions_inl.h.
+#include <ATen/CompositeImplicitAutogradFunctions_inl.h>

.venv/lib/python3.11/site-packages/torch/include/ATen/CompositeImplicitAutogradNestedTensorFunctions.h

@@ -0,0 +1,29 @@
+#include <ATen/core/TensorBody.h>
+
+// TODO Undo all logic introduced for Note [Avoiding Include Cycles In Static Dispatch]
+// Code introduced to avoid cyclic dependency in static dispatch is no longer
+// needed as static dispatch logic is moved from TensorBody.h, which caused cycles in the first place,
+// to Operators.cpp for supporting multiple backends with multiple kernels.
+//
+// Note [Avoiding Include Cycles In Static Dispatch]
+// In order to avoid #include cycles in the static dispatch build, we've carefully split out
+// the static function definition files into {DispatchKey}Functions.h and {DispatchKey}Functions_inl.h.
+//
+// Without this split, the include cycle looks like TensorBody.h -> CPUFunctions.h -> TensorBody.h.
+// - TensorBody.h #includes CPUFunctions.h in the static dispatch build, because the tensor methods
+//   all need to call into the fastpath C++ API defined in CPUFunctions.h. The methods are also all
+//   directly inlined into TensorBody.h.
+// - CPUFunctions.h #includes TensorBody.h because it contains function declarations for the entire C++ API,
+//   which include functions that have defaultable std::optional<Tensor> arguments.
+//   That requires knowing the full Tensor class definition.
+//
+// We break the cycle by doing the following:
+// - Split out CPUFunction.h into two files: CPUFunctions.h and CPUFunctions_inl.h
+// - CPUFunction.h is a dummy file that just includes the Tensor class and includes CPUFunctions_inl.,
+// - CPUFunctions_inl.h includes everything else
+// - (only in the static dispatch build) TensorBody.h makes sure to finish defining the Tensor class,
+//   and then it includes CPUFunctions_inl.h.
+// - All other files that want the cpu fastpath functions can include CPUFunctions.h directly.
+// - This also means that static dispatch build, CPUFunctions.h only needs to
+//   #include TensorBody.h, and it will automatically bring in CPUFunctions_inl.h.
+#include <ATen/CompositeImplicitAutogradNestedTensorFunctions_inl.h>

.venv/lib/python3.11/site-packages/torch/include/ATen/CompositeImplicitAutogradNestedTensorFunctions_inl.h

@@ -0,0 +1,25 @@
+#pragma once
+// @generated by torchgen/gen.py from DispatchKeyFunctions_inl.h
+
+// NB: The implementing C++ file is RegisterDispatchKey.cpp
+
+// The only #includes we need are for custom classes that have defaults in the C++ API
+#include <c10/core/MemoryFormat.h>
+#include <c10/core/Scalar.h>
+#include <ATen/core/Reduction.h>
+
+#if defined(AT_PER_OPERATOR_HEADERS) && defined(TORCH_ASSERT_ONLY_METHOD_OPERATORS)
+#error This change adds a dependency on all pytorch operators, meaning the     \
+  file will need to be re-compiled every time an operator is changed or added. \
+  Consider including a specific operator from                                  \
+  <ATen/ops/{my_operator}_compositeimplicitautogradnestedtensor_dispatch.h>.                   \
+  See NOTE [TORCH_ASSERT_ONLY_METHOD_OPERATORS].
+#endif
+
+#include <ATen/ops/randn_like_compositeimplicitautogradnestedtensor_dispatch.h>
+#include <ATen/ops/reshape_compositeimplicitautogradnestedtensor_dispatch.h>
+#include <ATen/ops/reshape_as_compositeimplicitautogradnestedtensor_dispatch.h>
+#include <ATen/ops/zeros_like_compositeimplicitautogradnestedtensor_dispatch.h>
+
+
+

.venv/lib/python3.11/site-packages/torch/include/ATen/Config.h

@@ -0,0 +1,21 @@
+#pragma once
+
+// Test these using #if AT_MKL_ENABLED(), not #ifdef, so that it's
+// obvious if you forgot to include Config.h
+//    c.f. https://stackoverflow.com/questions/33759787/generating-an-error-if-checked-boolean-macro-is-not-defined
+//
+// DO NOT put the macros for CUDA libraries in this file; they belong in cuda/CUDAConfig.h
+
+#define AT_MKLDNN_ENABLED() 1
+#define AT_MKLDNN_ACL_ENABLED() 0
+#define AT_MKL_ENABLED() 1
+#define AT_MKL_SEQUENTIAL() 0
+#define AT_POCKETFFT_ENABLED() 0
+#define AT_NNPACK_ENABLED() 1
+#define CAFFE2_STATIC_LINK_CUDA() 0
+#define AT_BUILD_WITH_BLAS() 1
+#define AT_BUILD_WITH_LAPACK() 1
+#define AT_PARALLEL_OPENMP 1
+#define AT_PARALLEL_NATIVE 0
+#define AT_BLAS_F2C() 0
+#define AT_BLAS_USE_CBLAS_DOT() 0

.venv/lib/python3.11/site-packages/torch/include/ATen/Context.h

@@ -0,0 +1,610 @@
+#pragma once
+
+#include <ATen/BlasBackend.h>
+#include <ATen/CPUGeneratorImpl.h>
+#include <ATen/DeviceAccelerator.h>
+#include <ATen/LinalgBackend.h>
+#include <ATen/core/ATenGeneral.h>
+#include <ATen/core/DeprecatedTypeProperties.h>
+#include <ATen/core/Generator.h>
+#include <ATen/core/LegacyTypeDispatch.h>
+#include <ATen/detail/AcceleratorHooksInterface.h>
+#include <ATen/detail/CUDAHooksInterface.h>
+#include <ATen/detail/HIPHooksInterface.h>
+#include <ATen/detail/IPUHooksInterface.h>
+#include <ATen/detail/MAIAHooksInterface.h>
+#include <ATen/detail/MPSHooksInterface.h>
+#include <ATen/detail/MTIAHooksInterface.h>
+#include <ATen/detail/PrivateUse1HooksInterface.h>
+#include <ATen/detail/XPUHooksInterface.h>
+#include <c10/core/QEngine.h>
+#include <c10/core/impl/DeviceGuardImplInterface.h>
+#include <c10/util/CallOnce.h>
+#include <c10/util/Exception.h>
+#include <c10/util/env.h>
+#include <c10/util/irange.h>
+
+#include <cstdint>
+#include <mutex>
+
+namespace at {
+
+class Tensor;
+
+enum class TORCH_API Float32MatmulPrecision { HIGHEST, HIGH, MEDIUM };
+
+class TORCH_API Context {
+ public:
+  Context();
+
+  const Generator& defaultGenerator(Device device) {
+    c10::DeviceType device_type = device.type();
+    initCUDAIfNeeded(device_type);
+    initHIPIfNeeded(device_type);
+    if (device_type == at::kCPU) {
+      return at::detail::getDefaultCPUGenerator();
+    } else if (device_type == at::kCUDA) {
+      return at::detail::getCUDAHooks().getDefaultCUDAGenerator(device.index());
+    } else if (device_type == at::kMPS) {
+      return at::detail::getMPSHooks().getDefaultMPSGenerator();
+    } else if (device_type == at::kXPU) {
+      return at::detail::getXPUHooks().getDefaultXPUGenerator(device.index());
+    } else if (device_type == at::kIPU) {
+      return at::detail::getIPUHooks().getDefaultIPUGenerator(device.index());
+    } else if (device_type == at::kPrivateUse1) {
+      return at::detail::getPrivateUse1Hooks().getDefaultGenerator(
+          device.index());
+    } else {
+      AT_ERROR(c10::DeviceTypeName(device_type), " device type not enabled.");
+    }
+  }
+  const AcceleratorHooksInterface& getAcceleratorHooksInterface(
+      std::optional<c10::DeviceType> opt_device_type = std::nullopt) {
+    c10::DeviceType device_type = opt_device_type.has_value()
+        ? opt_device_type.value()
+        : at::getAccelerator(true).value();
+    if (device_type == at::kCUDA) {
+      return at::detail::getCUDAHooks();
+    } else if (device_type == at::kXPU) {
+      return at::detail::getXPUHooks();
+    } else if (device_type == at::kMPS) {
+      return at::detail::getMPSHooks();
+    } else if (device_type == at::kPrivateUse1) {
+      return at::detail::getPrivateUse1Hooks();
+    } else if (device_type == at::kMTIA) {
+      return at::detail::getMTIAHooks();
+    } else if (device_type == at::kHIP) {
+      return at::detail::getHIPHooks();
+    } else {
+      AT_ERROR(
+          c10::DeviceTypeName(device_type), " device type not an accelerator.");
+    }
+  }
+  Device getDeviceFromPtr(void* data, c10::DeviceType device_type) {
+    initCUDAIfNeeded(device_type);
+    initHIPIfNeeded(device_type);
+    initXPUIfNeeded(device_type);
+    if (device_type == at::kCPU) {
+      return c10::DeviceType::CPU;
+    } else if (device_type == at::kCUDA) {
+      return at::detail::getCUDAHooks().getDeviceFromPtr(data);
+    } else if (device_type == at::kXPU) {
+      return at::detail::getXPUHooks().getDeviceFromPtr(data);
+    } else if (device_type == at::kPrivateUse1) {
+      return at::detail::getPrivateUse1Hooks().getDeviceFromPtr(data);
+    } else {
+      AT_ERROR(c10::DeviceTypeName(device_type), " device type not enabled.");
+    }
+  }
+  bool isPinnedPtr(
+      const void* data,
+      std::optional<c10::DeviceType> device_type = std::nullopt) {
+    auto opt_device_type =
+        device_type.has_value() ? device_type : at::getAccelerator();
+    if (!opt_device_type.has_value() || // there is no accelerator
+        !at::isAccelerator(
+            opt_device_type.value())) { // passed device not an accelerator
+      return false;
+    }
+    return getAcceleratorHooksInterface(opt_device_type.value())
+        .isPinnedPtr(data);
+  }
+  Allocator* getPinnedMemoryAllocator(
+      std::optional<c10::DeviceType> device_type = std::nullopt) {
+    return getAcceleratorHooksInterface(device_type).getPinnedMemoryAllocator();
+  }
+  static bool hasOpenMP();
+  static bool hasMKL();
+  static bool hasLAPACK();
+  static bool hasMKLDNN();
+  static bool hasMAGMA() {
+    return detail::getCUDAHooks().hasMAGMA();
+  }
+  static bool hasCUDA() {
+    return detail::getCUDAHooks().hasCUDA();
+  }
+  static bool hasMTIA() {
+    return detail::getMTIAHooks().hasMTIA();
+  }
+  static bool hasCUDART() {
+    return detail::getCUDAHooks().hasCUDART();
+  }
+  static long versionCUDART() {
+    return detail::getCUDAHooks().versionCUDART();
+  }
+  static bool hasCuDNN() {
+    return detail::getCUDAHooks().hasCuDNN();
+  }
+  static long versionCuDNN() {
+    return detail::getCUDAHooks().versionCuDNN();
+  }
+  static bool hasCuSOLVER() {
+    return detail::getCUDAHooks().hasCuSOLVER();
+  }
+  static bool hasCuBLASLt() {
+    return detail::getCUDAHooks().hasCuBLASLt();
+  }
+  static bool hasHIP() {
+    return detail::getHIPHooks().hasHIP();
+  }
+  static bool hasMPS() {
+    return detail::getMPSHooks().hasMPS();
+  }
+  static bool hasIPU() {
+    return c10::impl::hasDeviceGuardImpl(c10::DeviceType::IPU);
+  }
+  static bool hasXLA() {
+    return c10::impl::hasDeviceGuardImpl(c10::DeviceType::XLA);
+  }
+  static bool hasXPU() {
+    return detail::getXPUHooks().hasXPU();
+  }
+  static bool hasLazy() {
+    return c10::impl::hasDeviceGuardImpl(c10::DeviceType::Lazy);
+  }
+  static bool hasMAIA() {
+    return c10::impl::hasDeviceGuardImpl(c10::DeviceType::MAIA);
+  }
+  // defined in header so that getNonVariableType has ability to inline
+  // call_once check. getNonVariableType is called fairly frequently
+  void lazyInitCUDA() {
+    c10::call_once(thc_init, [&] { detail::getCUDAHooks().initCUDA(); });
+  }
+  void lazyInitHIP() {
+    c10::call_once(thh_init, [&] { detail::getHIPHooks().initHIP(); });
+  }
+  void lazyInitXPU() {
+    c10::call_once(thx_init, [&] { detail::getXPUHooks().initXPU(); });
+  }
+  void lazyInitMTIA() {
+    c10::call_once(th_mtia_init, [&] { detail::getMTIAHooks().initMTIA(); });
+  }
+  void lazyInitPrivateUse1() {
+    c10::call_once(thp_init, [&] {
+      if (isPrivateUse1HooksRegistered()) {
+        at::detail::getPrivateUse1Hooks().initPrivateUse1();
+      }
+    });
+  }
+  static const at::cuda::NVRTC& getNVRTC() {
+    return detail::getCUDAHooks().nvrtc();
+  }
+
+  static bool setFlushDenormal(bool on);
+
+  // NB: This method is *purely* whether or not a user requested
+  // that CuDNN was enabled, it doesn't actually say anything about
+  // whether or not CuDNN is actually usable.  Use cudnn_is_acceptable
+  // to test this instead
+  bool userEnabledCuDNN() const;
+  void setUserEnabledCuDNN(bool e);
+  bool userEnabledMkldnn() const;
+  void setUserEnabledMkldnn(bool e);
+  bool benchmarkCuDNN() const;
+  void setBenchmarkCuDNN(bool);
+  int benchmarkLimitCuDNN() const;
+  void setBenchmarkLimitCuDNN(int);
+  bool deterministicCuDNN() const;
+  void setDeterministicCuDNN(bool);
+  bool deterministicMkldnn() const;
+  void setDeterministicMkldnn(bool);
+  bool userEnabledNNPACK() const;
+  void setUserEnabledNNPACK(bool e);
+
+  // Note [Disabling Fused SDP Kernels]
+  // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  // Flash and Memory Efficient SDP kernels are enabled by default.
+  // However, they can be disabled by setting
+  // at::globalContext().setUserEnabledFlashSDP(false) flag.
+  // This is useful for debugging purposes. For example, if you want to
+  // compare the performance of the flash SDP kernels with the unfused
+  // kernel, you can disable the flash SDP kernels. By disabling
+  // the math SDP kernel, you can force your code to use flash kernels.
+  // The math SDP kernel can be disabled by setting
+  // at::globalContext().setUserEnabledMathSDP(false) flag.
+  void setSDPUseFlash(bool);
+  bool userEnabledFlashSDP() const;
+
+  void setSDPUseMemEfficient(bool);
+  bool userEnabledMemEfficientSDP() const;
+
+  void setSDPUseMath(bool);
+  bool userEnabledMathSDP() const;
+
+  void setSDPUseCuDNN(bool);
+  bool userEnabledCuDNNSDP() const;
+
+  void setAllowFP16BF16ReductionMathSDP(bool);
+  bool allowFP16BF16ReductionMathSDP() const;
+
+  void setSDPUseOverrideable(bool);
+  bool userEnabledOverrideableSDP() const;
+
+  at::LinalgBackend linalgPreferredBackend() const;
+  void setLinalgPreferredBackend(at::LinalgBackend);
+
+  at::BlasBackend blasPreferredBackend();
+  void setBlasPreferredBackend(at::BlasBackend);
+
+  // Note [Enabling Deterministic Operations]
+  // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  // Operations in PyTorch that normally act nondeterministically, but have an
+  // alternate deterministic implementation, should satisfy the following
+  // requirements:
+  //
+  // * Include this comment: "See Note [Enabling Deterministic Operations]"
+  //
+  // * Check the value of `at::globalContext().deterministicAlgorithms()` to
+  // toggle
+  //   between nondeterministic and deterministic implementations.
+  //
+  // * Have an entry in the list of PyTorch operations that toggle between
+  // nondeterministic
+  //   and deterministic implementations, in the docstring of
+  //   `use_deterministic_algorithms()` in torch/__init__.py
+  //
+  // `example_func()` below shows an example of toggling between
+  // nondeterministic and deterministic implementations:
+  //
+  //    void example_func() {
+  //      // See Note [Enabling Deterministic Operations]
+  //      if (at::globalContext().deterministicAlgorithms()) {
+  //        example_func_deterministic();
+  //      } else {
+  //        example_func_nondeterministic();
+  //      }
+  //    }
+
+  bool deterministicAlgorithms() const;
+  bool deterministicAlgorithmsWarnOnly() const;
+  void setDeterministicAlgorithms(bool, bool);
+  bool deterministicFillUninitializedMemory() const;
+  void setDeterministicFillUninitializedMemory(bool);
+
+  // Note [Writing Nondeterministic Operations]
+  // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  // Operations in PyTorch that act nondeterministically and do not have an
+  // alternate deterministic implementation should satisfy the following
+  // requirements:
+  //
+  // * Include this comment: "See Note [Writing Nondeterministic Operations]"
+  //
+  // * Include a comment explaining why the operation is nondeterministic.
+  //
+  // * Throw an error when `Context::deterministicAlgorithms()` is true. Most
+  //   of the time, this should be accomplished by calling
+  //   `at::globalContext().alertNotDeterminstic()`.  However, if the
+  //   nondeterministic behavior is caused by the CuBLAS workspace
+  //   configuration in CUDA >= 10.2,
+  //   `at::globalContext().alertCuBLASConfigNotDeterministic()` should be
+  //   called instead (in this case, a comment explaining why the operation is
+  //   nondeterministic is not necessary). See below for details on these
+  //   methods.
+  //
+  // * Have an entry in the list of nondeterministic PyTorch operations in the
+  //   docstring of `use_deterministic_algorithms()` in torch/__init__.py
+  //
+  // * Have a test function in `test/test_torch.py` whose name begins with
+  //   `test_nondeterministic_alert_`. Alternatively, if CuBLAS workspace
+  //   configuration is the reason for nondeterminism, the operation should be
+  //   included in the `test_cublas_config_nondeterministic_alert` test. Any new
+  //   tests should ideally follow a pattern similar to the existing ones.
+  //
+  // `example_func()` below shows an example of the comments and error-throwing
+  // code for a nondeterministic operation:
+  //
+  //    void example_func() {
+  //      // See Note [Writing Nondeterministic Operations]
+  //      // Nondeterministic because <reason>
+  //      at::globalContext().alertNondeterministic("example_func");
+  //      ...
+  //    }
+
+  // Throws an error if `Context::deterministicAlgorithms()` is true
+  static void alertNotDeterministic(c10::string_view const& caller);
+
+  // Throws an error if `Context::deterministicAlgorithms()` is true, CUDA
+  // >= 10.2, and CUBLAS_WORKSPACE_CONFIG is not set to either ":16:8" or
+  // ":4096:8". For more details:
+  // https://docs.nvidia.com/cuda/cublas/index.html#results-reproducibility
+  void alertCuBLASConfigNotDeterministic() const;
+
+  void setFloat32MatmulPrecision(const std::string& s);
+  bool allowTF32CuDNN() const;
+  void setAllowTF32CuDNN(bool);
+  bool allowTF32CuBLAS() const;
+  void setAllowTF32CuBLAS(bool);
+  Float32MatmulPrecision float32MatmulPrecision() const;
+  void setFloat32MatmulPrecision(Float32MatmulPrecision p);
+  bool allowFP16ReductionCuBLAS() const;
+  void setAllowFP16ReductionCuBLAS(bool);
+  bool allowBF16ReductionCuBLAS() const;
+  void setAllowBF16ReductionCuBLAS(bool);
+  at::QEngine qEngine() const;
+  void setQEngine(at::QEngine e);
+  static const std::vector<at::QEngine>& supportedQEngines();
+  static bool isXNNPACKAvailable();
+  void setCheckSparseTensorInvariants(bool e);
+  bool checkSparseTensorInvariants() const;
+  // This method is used to release the original weight after pre-packing.
+  // It should be called once before loading/running the model.
+  // NB: By default it is set to true for mobile builds.
+  void setReleaseWeightsWhenPrepacking(bool e);
+  bool releaseWeightsWhenPrepacking() const;
+
+  void setDisplayVmapFallbackWarnings(bool enabled);
+  bool areVmapFallbackWarningsEnabled() const;
+
+  void setDefaultMobileCPUAllocator();
+  void unsetDefaultMobileCPUAllocator();
+  bool allowFP16ReductionCPU() const;
+  void setAllowFP16ReductionCPU(bool);
+
+ private:
+  void initCUDAIfNeeded(c10::DeviceType p) {
+    if (p == c10::DeviceType::CUDA) {
+      lazyInitCUDA();
+    }
+  }
+  void initHIPIfNeeded(c10::DeviceType p) {
+    if (p == c10::DeviceType::HIP) {
+      lazyInitHIP();
+    }
+  }
+  void initXPUIfNeeded(c10::DeviceType p) {
+    if (p == c10::DeviceType::XPU) {
+      lazyInitXPU();
+    }
+  }
+  static bool checkCuBLASConfigDeterministic();
+  c10::once_flag thc_init;
+  c10::once_flag thh_init;
+  c10::once_flag thx_init;
+  c10::once_flag th_mtia_init;
+  c10::once_flag thp_init;
+  bool enabled_cudnn = true;
+  bool deterministic_cudnn = false;
+  bool deterministic_mkldnn = false;
+  bool _deterministic_algorithms = false;
+  bool _deterministic_algorithms_warn_only = false;
+  bool _deterministic_fill_uninitialized_memory = true;
+  bool enabled_flashSDP = true;
+  bool enabled_mem_efficientSDP = true;
+  bool enabled_mathSDP = true;
+  bool enabled_cudnnSDP = true;
+  bool enabled_overrideable = true;
+  bool allow_fp16_bf16_reduction_mathSDP = false;
+#ifdef USE_ROCM
+  bool benchmark_cudnn = true;
+#else
+  bool benchmark_cudnn = false;
+#endif
+  Float32MatmulPrecision float32_matmul_precision =
+      c10::utils::check_env("TORCH_ALLOW_TF32_CUBLAS_OVERRIDE") == true
+      ? at::Float32MatmulPrecision::HIGH
+      : at::Float32MatmulPrecision::HIGHEST;
+  int benchmark_limit_cudnn = 10;
+  bool allow_tf32_cudnn = true;
+  bool allow_fp16_reduction_cublas = true;
+  bool allow_bf16_reduction_cublas = true;
+  bool enabled_mkldnn = true;
+  bool enabled_nnpack = true;
+  at::LinalgBackend linalg_preferred_backend =
+      c10::utils::check_env("TORCH_LINALG_PREFER_CUSOLVER") == true
+      ? at::LinalgBackend::Cusolver
+      : at::LinalgBackend::Default;
+  at::BlasBackend blas_preferred_backend =
+#ifdef USE_ROCM
+      (c10::utils::check_env("TORCH_BLAS_PREFER_HIPBLASLT") != false)
+#else
+      (c10::utils::check_env("TORCH_BLAS_PREFER_CUBLASLT") == true)
+#endif
+      ? at::BlasBackend::Cublaslt
+      : at::BlasBackend::Cublas;
+#ifdef C10_MOBILE
+  bool release_original_weights = true;
+#else
+  bool release_original_weights = false;
+#endif
+  bool display_vmap_fallback_warnings_ = false;
+  std::optional<at::QEngine> quantized_engine = std::nullopt;
+  bool enable_sparse_tensor_invariant_checks = false;
+  bool allow_fp16_reduction_cpu = false;
+
+  Allocator* prev_allocator_ptr_{nullptr};
+};
+
+TORCH_API Context& globalContext();
+
+inline void init() {
+  globalContext();
+}
+
+TORCH_API Allocator* getCPUAllocator();
+
+inline DeprecatedTypeProperties& getDeprecatedTypeProperties(
+    Backend p,
+    ScalarType s) {
+  return globalDeprecatedTypePropertiesRegistry().getDeprecatedTypeProperties(
+      p, s);
+}
+
+inline DeprecatedTypeProperties& CPU(ScalarType s) {
+  return globalDeprecatedTypePropertiesRegistry().getDeprecatedTypeProperties(
+      Backend::CPU, s);
+}
+
+inline DeprecatedTypeProperties& CUDA(ScalarType s) {
+  return globalDeprecatedTypePropertiesRegistry().getDeprecatedTypeProperties(
+      Backend::CUDA, s);
+}
+
+inline DeprecatedTypeProperties& HIP(ScalarType s) {
+  return globalDeprecatedTypePropertiesRegistry().getDeprecatedTypeProperties(
+      Backend::HIP, s);
+}
+
+inline DeprecatedTypeProperties& MPS(ScalarType s) {
+  return globalDeprecatedTypePropertiesRegistry().getDeprecatedTypeProperties(
+      Backend::MPS, s);
+}
+
+inline bool hasCUDA() {
+  return globalContext().hasCUDA();
+}
+
+inline bool hasMTIA() {
+  return globalContext().hasMTIA();
+}
+
+inline bool hasHIP() {
+  return globalContext().hasHIP();
+}
+
+inline bool hasIPU() {
+  return globalContext().hasIPU();
+}
+
+inline bool hasXLA() {
+  return globalContext().hasXLA();
+}
+
+inline bool hasMPS() {
+  return globalContext().hasMPS();
+}
+
+inline bool hasMAIA() {
+  return globalContext().hasMAIA();
+}
+
+inline bool hasXPU() {
+  return globalContext().hasXPU();
+}
+
+// Despite its name, this function returns the number of *CUDA* GPUs.
+inline size_t getNumGPUs() {
+  // WARNING: DO NOT ADD LOGIC TO HANDLE OTHER DEVICE TYPES TO THIS
+  // FUNCTION.  If you are interested in interrogating the number of
+  // devices for a specific device type, add that function to the
+  // relevant library (e.g., similar to at::cuda::device_count())
+  if (hasCUDA() && hasHIP()) {
+    throw std::runtime_error(
+        "Enabling both CUDA and HIP in ATen is not supported, as HIP masquerades "
+        "to be CUDA (e.g., when you say CUDA, on a HIP build of ATen, this actually "
+        "means HIP.  Rebuild PyTorch with one or the other disabled.");
+  } else if (hasCUDA()) {
+    return detail::getCUDAHooks().getNumGPUs();
+  } else if (hasHIP()) {
+    return detail::getHIPHooks().getNumGPUs();
+  } else {
+    return 0;
+  }
+}
+
+inline bool hasOpenMP() {
+  return globalContext().hasOpenMP();
+}
+
+inline bool hasMKL() {
+  return globalContext().hasMKL();
+}
+
+inline bool hasLAPACK() {
+  return globalContext().hasLAPACK();
+}
+
+inline bool hasMAGMA() {
+  return globalContext().hasMAGMA();
+}
+
+inline bool hasMKLDNN() {
+  return globalContext().hasMKLDNN();
+}
+
+inline void manual_seed(uint64_t seed) {
+  auto gen = globalContext().defaultGenerator(c10::DeviceType::CPU);
+  {
+    // See Note [Acquire lock when using random generators]
+    std::lock_guard<std::mutex> lock(gen.mutex());
+    gen.set_current_seed(seed);
+  }
+  // NB: Sometimes we build with CUDA, but we don't have any GPUs
+  // available. In that case, we must not seed CUDA; it will fail!
+  const auto cuda_num_gpus = detail::getCUDAHooks().getNumGPUs();
+  if (hasCUDA() && cuda_num_gpus > 0) {
+    for (const auto i : c10::irange(cuda_num_gpus)) {
+      auto cuda_gen = globalContext().defaultGenerator(
+          Device(at::kCUDA, static_cast<c10::DeviceIndex>(i)));
+      {
+        // See Note [Acquire lock when using random generators]
+        std::lock_guard<std::mutex> lock(cuda_gen.mutex());
+        cuda_gen.set_current_seed(seed);
+      }
+    }
+  }
+
+  const auto xpu_num_gpus = detail::getXPUHooks().getNumGPUs();
+  if (hasXPU() && xpu_num_gpus) {
+    for (const auto i : c10::irange(xpu_num_gpus)) {
+      auto xpu_gen = globalContext().defaultGenerator(
+          Device(at::kXPU, static_cast<c10::DeviceIndex>(i)));
+      {
+        // See Note [Acquire lock when using random generators]
+        std::lock_guard<std::mutex> lock(xpu_gen.mutex());
+        xpu_gen.set_current_seed(seed);
+      }
+    }
+  }
+
+  if (hasMPS()) {
+    auto mps_gen = globalContext().defaultGenerator(c10::DeviceType::MPS);
+    // See Note [Acquire lock when using random generators]
+    std::lock_guard<std::mutex> lock(mps_gen.mutex());
+    mps_gen.set_current_seed(seed);
+  }
+}
+
+// When the global flag `allow_tf32` is set to true, cuBLAS handles are
+// automatically configured to use math mode CUBLAS_TF32_TENSOR_OP_MATH.
+// For some operators, such as addmv, TF32 offers no performance improvement
+// but causes precision loss. To help this case, this class implements
+// a RAII guard that can be used to quickly disable TF32 within its scope.
+//
+// Usage:
+//     NoTF32Guard disable_tf32;
+struct TORCH_API NoTF32Guard {
+  NoTF32Guard();
+  ~NoTF32Guard();
+  static bool should_disable_tf32();
+
+ private:
+  bool changed = false;
+};
+
+struct TORCH_API ROCmBackwardPassGuard {
+  ROCmBackwardPassGuard();
+  ~ROCmBackwardPassGuard();
+  static bool is_backward_pass();
+};
+
+} // namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/Device.h

@@ -0,0 +1,2 @@
+#pragma once
+#include <c10/core/Device.h>

.venv/lib/python3.11/site-packages/torch/include/ATen/DeviceAccelerator.h

@@ -0,0 +1,27 @@
+#pragma once
+
+#include <c10/core/DeviceType.h>
+#include <c10/macros/Macros.h>
+
+#include <ATen/detail/MTIAHooksInterface.h>
+#include <optional>
+
+// This file defines the top level Accelerator concept for PyTorch.
+// A device is an accelerator per the definition here if:
+// - It is mutually exclusive with all other accelerators
+// - It performs asynchronous compute via a Stream/Event system
+// - It provides a set of common APIs as defined by AcceleratorHooksInterface
+//
+// As of today, accelerator devices are (in no particular order):
+// CUDA, MTIA, XPU, HIP, MPS, PrivateUse1
+
+namespace at {
+
+// Ensures that only one accelerator is available (at
+// compile time if possible) and return it.
+// When checked is true, the returned optional always has a value.
+TORCH_API std::optional<c10::DeviceType> getAccelerator(bool checked = false);
+
+TORCH_API bool isAccelerator(c10::DeviceType d);
+
+} // namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/Dispatch.h

@@ -0,0 +1,808 @@
+#pragma once
+
+#include <ATen/core/DeprecatedTypeProperties.h>
+#include <c10/macros/Macros.h>
+#include <c10/util/Exception.h>
+#include <c10/util/Half.h>
+#include <c10/util/Metaprogramming.h>
+#include <c10/util/complex.h>
+#include <c10/util/string_view.h>
+
+#ifdef __CUDACC__
+#include <cuda.h> // For CUDA_VERSION
+#endif
+
+#ifdef TEMPLATE_SELECTIVE_BUILD
+#include <ATen/selected_mobile_ops.h>
+#else
+namespace at {
+/**
+ * The method should_include_kernel_dtype() returns true/false
+ * based on whether the switching code for a specific dtype should be
+ * included based on build time constants generated from tracing model
+ * execution. This method will be implemented via code-generation and
+ * included in this file when code-gen is ready.
+ */
+inline constexpr bool should_include_kernel_dtype(
+    const char* /*kernel_tag_str*/,
+    at::ScalarType /*scalar_type*/
+) {
+  return true;
+}
+} // namespace at
+#endif
+
+/**
+ * In the Facebook internal build (using BUCK), this macro is enabled by
+ * passing in -c pt.enable_record_kernel_dtype=1 when building the tracer
+ * binary.
+ */
+#if defined ENABLE_RECORD_KERNEL_FUNCTION_DTYPE
+namespace at {
+namespace detail {
+TORCH_API void record_kernel_function_dtype(std::string name);
+}
+} // namespace at
+
+#define RECORD_KERNEL_FUNCTION_DTYPE(NAME, enum_type) \
+  at::detail::record_kernel_function_dtype(           \
+      std::string(NAME) + "$" + toString(enum_type));
+#else
+#define RECORD_KERNEL_FUNCTION_DTYPE(NAME, enum_type)
+#endif
+
+#define AT_PRIVATE_CHECK_SELECTIVE_BUILD(enum_type)   \
+  do {                                                \
+    if constexpr (!at::should_include_kernel_dtype(   \
+                      at_dispatch_name, enum_type)) { \
+      AT_ERROR(                                       \
+          "dtype '",                                  \
+          toString(enum_type),                        \
+          "' not selected for kernel tag ",           \
+          at_dispatch_name);                          \
+    }                                                 \
+  } while (0)
+
+#define AT_PRIVATE_CASE_TYPE_USING_HINT(enum_type, HINT, ...)           \
+  case enum_type: {                                                     \
+    AT_PRIVATE_CHECK_SELECTIVE_BUILD(enum_type);                        \
+    using HINT C10_UNUSED = c10::impl::ScalarTypeToCPPTypeT<enum_type>; \
+    return __VA_ARGS__();                                               \
+  }
+
+#define AT_DISPATCH_CASE(enum_type, ...) \
+  AT_PRIVATE_CASE_TYPE_USING_HINT(enum_type, scalar_t, __VA_ARGS__)
+
+#define AT_DISPATCH_CASE_QINT(enum_type, scalar_type, ...)            \
+  case enum_type: {                                                   \
+    AT_PRIVATE_CHECK_SELECTIVE_BUILD(enum_type);                      \
+    using scalar_t = scalar_type;                                     \
+    using underlying_t C10_UNUSED = typename scalar_t::underlying;    \
+    const auto& SCALAR_TYPE C10_UNUSED = enum_type;                   \
+    const auto& UNDERLYING_TYPE C10_UNUSED = toUnderlying(enum_type); \
+    return __VA_ARGS__();                                             \
+  }
+
+#define AT_QINT_SUB_BYTE_PRIVATE_CASE_TYPE(                           \
+    enum_type, scalar_type, bitwidth, qmin, qmax, ...)                \
+  case enum_type: {                                                   \
+    AT_PRIVATE_CHECK_SELECTIVE_BUILD(enum_type);                      \
+    using scalar_t = scalar_type;                                     \
+    using underlying_t C10_UNUSED = typename scalar_t::underlying;    \
+    const auto& SCALAR_TYPE C10_UNUSED = enum_type;                   \
+    const auto& UNDERLYING_TYPE C10_UNUSED = toUnderlying(enum_type); \
+    C10_UNUSED int bit_width = bitwidth;                              \
+    C10_UNUSED int64_t quant_min = qmin;                              \
+    C10_UNUSED int64_t quant_max = qmax;                              \
+    return __VA_ARGS__();                                             \
+  }
+
+namespace detail {
+
+inline at::ScalarType scalar_type(at::ScalarType s) {
+  return s;
+}
+
+C10_DEPRECATED_MESSAGE(
+    "passing at::DeprecatedTypeProperties to an AT_DISPATCH macro is deprecated, "
+    "pass an at::ScalarType instead")
+inline at::ScalarType scalar_type(const at::DeprecatedTypeProperties& t) {
+  return t.scalarType();
+}
+
+C10_DEPRECATED_MESSAGE(
+    "AT_DISPATCH_ALL_TYPES_AND_HALF is deprecated, "
+    "use AT_DISPATCH_ALL_TYPES_AND(at::ScalarType::Half, ...) instead")
+inline void deprecated_AT_DISPATCH_ALL_TYPES_AND_HALF() {}
+
+C10_DEPRECATED_MESSAGE(
+    "AT_DISPATCH_ALL_TYPES_AND_HALF_AND_COMPLEX is deprecated, "
+    "use AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND(at::ScalarType::Half, ...) "
+    "instead")
+inline void deprecated_AT_DISPATCH_ALL_TYPES_AND_HALF_AND_COMPLEX() {}
+
+} // namespace detail
+
+// The AT_DISPATCH_* family of macros provides the ability to
+// conveniently generate specializations of a kernel over all of the
+// dtypes we care about in PyTorch.  We call it "dispatch" because
+// we are "dispatching" to the correct, dtype-specific kernel.
+//
+// A standard usage looks like:
+//
+//      AT_DISPATCH_ALL_TYPES(self.scalar_type(), "op_name", [&] {
+//          // Your code here, with 'scalar_t' now defined to
+//          // be the dtype in question
+//      });
+//
+// There are many variations of this macro, so it's important to
+// understand exactly /which/ dtypes you want to get instantiated, as
+// well as what the "default" set is.
+//
+// The default set of dtypes that are instantiated (e.g., by
+// AT_DISPATCH_ALL_TYPES) are floating point types (float, double),
+// and integral types (int32_t, int64_t, int16_t, int8_t, uint8_t),
+// but NOT booleans (bool), half-precision floats (Half) or
+// complex number (c10::complex<float>, c10::complex<double>).
+// This "cut" is somewhat historical (the default types are the
+// ones that TH historically supported), but it also reflects the
+// fact that the non-default types are "poorly" behaved (booleans
+// are NOT integers mod 2, half precision operations ~essentially
+// don't exist on CPU, complex numbers are an experimental application).
+//
+// Here are the questions you should generally ask to decide which
+// dispatch you want:
+//
+// 1. Is this an integral or floating point specific operation?
+//    (If so, you'll want one of the FLOATING or INTEGRAL macros.)
+//
+// 2. Should half be supported?  (If you're on CPU, the answer is almost
+//    definitely no.  If you do want support, use one of the AND_HALF
+//    macros)
+//
+// Much rarer situations:
+//
+// 3. Should bool be supported?  (You often have to write your kernel
+//    differently if arithmetic operations are involved.)  If so,
+//    Use AT_DISPATCH_ALL_TYPES_AND along with ScalarType::Bool
+//
+// 4. Should complex be supported?  The answer is almost always no,
+//    unless you are working on "generic" code that should work on
+//    all dtypes.
+//
+// Parameters:
+// -----------
+//
+// 1. The NAME argument is a "tag" that is used to trace and then
+//    conditionally compile fragments of the case statements such
+//    that the kernel functions are specialized only for the dtypes
+//    that are needed. The NAME parameter *must* be a build time
+//    const char* (can't be std::string, etc...)
+//
+// Please ensure that the NAME is unique for every implementation
+// or you run the risk of over-including code for the kernel
+// functions. There is no risk of missing out on any code, so
+// it's mostly a risk of a Type-2 error, and not a Type-1 error.
+//
+// Switch-like syntax:
+// -------------------
+// There is also a switch-case like syntax which is useful if a kernel
+// needs to be specialized for particular scalar types
+//
+//      AT_DISPATCH_SWITCH(self.scalar_type(), "op_name",
+//          AT_DISPATCH_CASE_INTEGRAL_TYPES([&] {
+//            op_integral<scalar_t>(iter);
+//          })
+//          AT_DISPATCH_CASE_FLOATING_TYPES([&] {
+//            op_floating<scalar_t>(iter);
+//          })
+//          AT_DISPATCH_CASE(kBool, [&] {
+//            op_bool(iter);
+//          })
+//      );
+//
+// For each AT_DISPATCH_FOO macro, there is a corresponding
+// AT_DISPATCH_CASE_FOO macro which can be used inside of an
+// AT_DISPATCH_SWITCH block.
+
+// NB: the the_type variable is not used, but we have kept it for
+// backwards compatibility.  It's probably not used by anyone though;
+// but we're just being safe (and it doesn't hurt.)  Note we must
+// use it to shut up warnings about unused store.
+
+#define AT_DISPATCH_SWITCH(TYPE, NAME, ...)                                 \
+  [&] {                                                                     \
+    const auto& the_type = TYPE;                                            \
+    constexpr const char* at_dispatch_name = NAME;                          \
+    /* don't use TYPE again in case it is an expensive or side-effect op */ \
+    at::ScalarType _st = ::detail::scalar_type(the_type);                   \
+    RECORD_KERNEL_FUNCTION_DTYPE(at_dispatch_name, _st);                    \
+    switch (_st) {                                                          \
+      __VA_ARGS__                                                           \
+      default:                                                              \
+        AT_ERROR(                                                           \
+            '"',                                                            \
+            at_dispatch_name,                                               \
+            "\" not implemented for '",                                     \
+            toString(_st),                                                  \
+            "'");                                                           \
+    }                                                                       \
+  }()
+
+#define AT_DISPATCH_CASE_FLOATING_TYPES(...)            \
+  AT_DISPATCH_CASE(at::ScalarType::Double, __VA_ARGS__) \
+  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)
+
+#define AT_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(TYPE, NAME, AT_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
+
+#define AT_DISPATCH_CASE_FLOATING_TYPES_AND_HALF(...)   \
+  AT_DISPATCH_CASE(at::ScalarType::Double, __VA_ARGS__) \
+  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)  \
+  AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)
+
+#define AT_DISPATCH_FLOATING_TYPES_AND_HALF(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                        \
+      TYPE, NAME, AT_DISPATCH_CASE_FLOATING_TYPES_AND_HALF(__VA_ARGS__))
+
+#define AT_DISPATCH_CASE_REDUCED_FLOATING_TYPES(...)  \
+  AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__) \
+  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)
+
+#define AT_DISPATCH_REDUCED_FLOATING_TYPES(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                       \
+      TYPE, NAME, AT_DISPATCH_CASE_REDUCED_FLOATING_TYPES(__VA_ARGS__))
+
+#define AT_DISPATCH_CASE_FLOATING_TYPES_AND(SCALARTYPE, ...) \
+  AT_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__)               \
+  AT_DISPATCH_CASE(SCALARTYPE, __VA_ARGS__)
+
+#define AT_DISPATCH_FLOATING_TYPES_AND(SCALARTYPE, TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                               \
+      TYPE,                                                         \
+      NAME,                                                         \
+      AT_DISPATCH_CASE_FLOATING_TYPES_AND(SCALARTYPE, __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_FLOATING_TYPES_AND2(SCALARTYPE1, SCALARTYPE2, ...) \
+  AT_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__)                              \
+  AT_DISPATCH_CASE(SCALARTYPE1, __VA_ARGS__)                                \
+  AT_DISPATCH_CASE(SCALARTYPE2, __VA_ARGS__)
+
+#define AT_DISPATCH_FLOATING_TYPES_AND2(       \
+    SCALARTYPE1, SCALARTYPE2, TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                          \
+      TYPE,                                    \
+      NAME,                                    \
+      AT_DISPATCH_CASE_FLOATING_TYPES_AND2(    \
+          SCALARTYPE1, SCALARTYPE2, __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_FLOATING_TYPES_AND3(   \
+    SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, ...) \
+  AT_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__)  \
+  AT_DISPATCH_CASE(SCALARTYPE1, __VA_ARGS__)    \
+  AT_DISPATCH_CASE(SCALARTYPE2, __VA_ARGS__)    \
+  AT_DISPATCH_CASE(SCALARTYPE3, __VA_ARGS__)
+
+#define AT_DISPATCH_FLOATING_TYPES_AND3(                    \
+    SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                       \
+      TYPE,                                                 \
+      NAME,                                                 \
+      AT_DISPATCH_CASE_FLOATING_TYPES_AND3(                 \
+          SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_FLOATING_TYPES_AND4(                \
+    SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, SCALARTYPE4, ...) \
+  AT_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__)               \
+  AT_DISPATCH_CASE(SCALARTYPE1, __VA_ARGS__)                 \
+  AT_DISPATCH_CASE(SCALARTYPE2, __VA_ARGS__)                 \
+  AT_DISPATCH_CASE(SCALARTYPE3, __VA_ARGS__)                 \
+  AT_DISPATCH_CASE(SCALARTYPE4, __VA_ARGS__)
+
+#define AT_DISPATCH_FLOATING_TYPES_AND4(                                 \
+    SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, SCALARTYPE4, TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                                    \
+      TYPE,                                                              \
+      NAME,                                                              \
+      AT_DISPATCH_CASE_FLOATING_TYPES_AND4(                              \
+          SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, SCALARTYPE4, __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_COMPLEX_TYPES(...)                    \
+  AT_DISPATCH_CASE(at::ScalarType::ComplexDouble, __VA_ARGS__) \
+  AT_DISPATCH_CASE(at::ScalarType::ComplexFloat, __VA_ARGS__)
+
+#define AT_DISPATCH_COMPLEX_TYPES(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(TYPE, NAME, AT_DISPATCH_CASE_COMPLEX_TYPES(__VA_ARGS__))
+
+#define AT_DISPATCH_CASE_COMPLEX_TYPES_AND(SCALARTYPE, ...) \
+  AT_DISPATCH_CASE_COMPLEX_TYPES(__VA_ARGS__)               \
+  AT_DISPATCH_CASE(SCALARTYPE, __VA_ARGS__)
+
+#define AT_DISPATCH_COMPLEX_TYPES_AND(SCALARTYPE, TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                              \
+      TYPE, NAME, AT_DISPATCH_CASE_COMPLEX_TYPES_AND(SCALARTYPE, __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_FLOATING_AND_COMPLEX_TYPES(...) \
+  AT_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__)           \
+  AT_DISPATCH_CASE_COMPLEX_TYPES(__VA_ARGS__)
+
+#define AT_DISPATCH_FLOATING_AND_COMPLEX_TYPES(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                           \
+      TYPE, NAME, AT_DISPATCH_CASE_FLOATING_AND_COMPLEX_TYPES(__VA_ARGS__))
+
+#define AT_DISPATCH_CASE_FLOATING_AND_COMPLEX_TYPES_AND1(SCALARTYPE, ...) \
+  AT_DISPATCH_CASE_FLOATING_AND_COMPLEX_TYPES(__VA_ARGS__)                \
+  AT_DISPATCH_CASE(SCALARTYPE, __VA_ARGS__)
+
+#define AT_DISPATCH_FLOATING_AND_COMPLEX_TYPES_AND1(    \
+    SCALARTYPE, TYPE, NAME, ...)                        \
+  AT_DISPATCH_SWITCH(                                   \
+      TYPE,                                             \
+      NAME,                                             \
+      AT_DISPATCH_CASE_FLOATING_AND_COMPLEX_TYPES_AND1( \
+          SCALARTYPE, __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_FLOATING_AND_COMPLEX_TYPES_AND2(  \
+    SCALARTYPE1, SCALARTYPE2, ...)                         \
+  AT_DISPATCH_CASE_FLOATING_AND_COMPLEX_TYPES(__VA_ARGS__) \
+  AT_DISPATCH_CASE(SCALARTYPE1, __VA_ARGS__)               \
+  AT_DISPATCH_CASE(SCALARTYPE2, __VA_ARGS__)
+
+#define AT_DISPATCH_FLOATING_AND_COMPLEX_TYPES_AND2(    \
+    SCALARTYPE1, SCALARTYPE2, TYPE, NAME, ...)          \
+  AT_DISPATCH_SWITCH(                                   \
+      TYPE,                                             \
+      NAME,                                             \
+      AT_DISPATCH_CASE_FLOATING_AND_COMPLEX_TYPES_AND2( \
+          SCALARTYPE1, SCALARTYPE2, __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_FLOATING_AND_COMPLEX_TYPES_AND3(  \
+    SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, ...)            \
+  AT_DISPATCH_CASE_FLOATING_AND_COMPLEX_TYPES(__VA_ARGS__) \
+  AT_DISPATCH_CASE(SCALARTYPE1, __VA_ARGS__)               \
+  AT_DISPATCH_CASE(SCALARTYPE2, __VA_ARGS__)               \
+  AT_DISPATCH_CASE(SCALARTYPE3, __VA_ARGS__)
+
+#define AT_DISPATCH_FLOATING_AND_COMPLEX_TYPES_AND3(        \
+    SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                       \
+      TYPE,                                                 \
+      NAME,                                                 \
+      AT_DISPATCH_CASE_FLOATING_AND_COMPLEX_TYPES_AND3(     \
+          SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_FLOATING_AND_COMPLEX_TYPES_AND4(    \
+    SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, SCALARTYPE4, ...) \
+  AT_DISPATCH_CASE_FLOATING_AND_COMPLEX_TYPES(__VA_ARGS__)   \
+  AT_DISPATCH_CASE(SCALARTYPE1, __VA_ARGS__)                 \
+  AT_DISPATCH_CASE(SCALARTYPE2, __VA_ARGS__)                 \
+  AT_DISPATCH_CASE(SCALARTYPE3, __VA_ARGS__)                 \
+  AT_DISPATCH_CASE(SCALARTYPE4, __VA_ARGS__)
+
+#define AT_DISPATCH_FLOATING_AND_COMPLEX_TYPES_AND4(                     \
+    SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, SCALARTYPE4, TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                                    \
+      TYPE,                                                              \
+      NAME,                                                              \
+      AT_DISPATCH_CASE_FLOATING_AND_COMPLEX_TYPES_AND4(                  \
+          SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, SCALARTYPE4, __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_FLOATING_AND_COMPLEX_TYPES_AND5(                 \
+    SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, SCALARTYPE4, SCALARTYPE5, ...) \
+  AT_DISPATCH_CASE_FLOATING_AND_COMPLEX_TYPES(__VA_ARGS__)                \
+  AT_DISPATCH_CASE(SCALARTYPE1, __VA_ARGS__)                              \
+  AT_DISPATCH_CASE(SCALARTYPE2, __VA_ARGS__)                              \
+  AT_DISPATCH_CASE(SCALARTYPE3, __VA_ARGS__)                              \
+  AT_DISPATCH_CASE(SCALARTYPE4, __VA_ARGS__)                              \
+  AT_DISPATCH_CASE(SCALARTYPE5, __VA_ARGS__)
+
+#define AT_DISPATCH_FLOATING_AND_COMPLEX_TYPES_AND5(    \
+    SCALARTYPE1,                                        \
+    SCALARTYPE2,                                        \
+    SCALARTYPE3,                                        \
+    SCALARTYPE4,                                        \
+    SCALARTYPE5,                                        \
+    TYPE,                                               \
+    NAME,                                               \
+    ...)                                                \
+  AT_DISPATCH_SWITCH(                                   \
+      TYPE,                                             \
+      NAME,                                             \
+      AT_DISPATCH_CASE_FLOATING_AND_COMPLEX_TYPES_AND5( \
+          SCALARTYPE1,                                  \
+          SCALARTYPE2,                                  \
+          SCALARTYPE3,                                  \
+          SCALARTYPE4,                                  \
+          SCALARTYPE5,                                  \
+          __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_FLOATING_AND_COMPLEX_TYPES_AND6(  \
+    SCALARTYPE1,                                           \
+    SCALARTYPE2,                                           \
+    SCALARTYPE3,                                           \
+    SCALARTYPE4,                                           \
+    SCALARTYPE5,                                           \
+    SCALARTYPE6,                                           \
+    ...)                                                   \
+  AT_DISPATCH_CASE_FLOATING_AND_COMPLEX_TYPES(__VA_ARGS__) \
+  AT_DISPATCH_CASE(SCALARTYPE1, __VA_ARGS__)               \
+  AT_DISPATCH_CASE(SCALARTYPE2, __VA_ARGS__)               \
+  AT_DISPATCH_CASE(SCALARTYPE3, __VA_ARGS__)               \
+  AT_DISPATCH_CASE(SCALARTYPE4, __VA_ARGS__)               \
+  AT_DISPATCH_CASE(SCALARTYPE5, __VA_ARGS__)               \
+  AT_DISPATCH_CASE(SCALARTYPE6, __VA_ARGS__)
+
+#define AT_DISPATCH_FLOATING_AND_COMPLEX_TYPES_AND6(    \
+    SCALARTYPE1,                                        \
+    SCALARTYPE2,                                        \
+    SCALARTYPE3,                                        \
+    SCALARTYPE4,                                        \
+    SCALARTYPE5,                                        \
+    SCALARTYPE6,                                        \
+    TYPE,                                               \
+    NAME,                                               \
+    ...)                                                \
+  AT_DISPATCH_SWITCH(                                   \
+      TYPE,                                             \
+      NAME,                                             \
+      AT_DISPATCH_CASE_FLOATING_AND_COMPLEX_TYPES_AND6( \
+          SCALARTYPE1,                                  \
+          SCALARTYPE2,                                  \
+          SCALARTYPE3,                                  \
+          SCALARTYPE4,                                  \
+          SCALARTYPE5,                                  \
+          SCALARTYPE6,                                  \
+          __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_INTEGRAL_TYPES(...)          \
+  AT_DISPATCH_CASE(at::ScalarType::Byte, __VA_ARGS__) \
+  AT_DISPATCH_CASE(at::ScalarType::Char, __VA_ARGS__) \
+  AT_DISPATCH_CASE(at::ScalarType::Int, __VA_ARGS__)  \
+  AT_DISPATCH_CASE(at::ScalarType::Long, __VA_ARGS__) \
+  AT_DISPATCH_CASE(at::ScalarType::Short, __VA_ARGS__)
+
+#define AT_DISPATCH_INTEGRAL_TYPES(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(TYPE, NAME, AT_DISPATCH_CASE_INTEGRAL_TYPES(__VA_ARGS__))
+
+#define AT_DISPATCH_CASE_INTEGRAL_TYPES_AND(SCALARTYPE, ...) \
+  AT_DISPATCH_CASE_INTEGRAL_TYPES(__VA_ARGS__)               \
+  AT_DISPATCH_CASE(SCALARTYPE, __VA_ARGS__)
+
+#define AT_DISPATCH_INTEGRAL_TYPES_AND(SCALARTYPE, TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                               \
+      TYPE,                                                         \
+      NAME,                                                         \
+      AT_DISPATCH_CASE_INTEGRAL_TYPES_AND(SCALARTYPE, __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_ALL_TYPES(...)        \
+  AT_DISPATCH_CASE_INTEGRAL_TYPES(__VA_ARGS__) \
+  AT_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__)
+
+#define AT_DISPATCH_ALL_TYPES(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(TYPE, NAME, AT_DISPATCH_CASE_ALL_TYPES(__VA_ARGS__))
+
+#define AT_DISPATCH_CASE_QINT_TYPES(...)                      \
+  AT_DISPATCH_CASE_QINT(at::kQInt8, at::qint8, __VA_ARGS__)   \
+  AT_DISPATCH_CASE_QINT(at::kQUInt8, at::quint8, __VA_ARGS__) \
+  AT_DISPATCH_CASE_QINT(at::kQInt32, at::qint32, __VA_ARGS__)
+
+#define AT_DISPATCH_QINT_TYPES(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(TYPE, NAME, AT_DISPATCH_CASE_QINT_TYPES(__VA_ARGS__))
+
+#define AT_DISPATCH_CASE_QINT_TYPES_AND(SCALARTYPE, ...) \
+  AT_DISPATCH_CASE_QINT_TYPES(__VA_ARGS__)               \
+  AT_DISPATCH_CASE(SCALARTYPE, __VA_ARGS__)
+
+#define AT_DISPATCH_QINT_TYPES_AND(SCALARTYPE, TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                           \
+      TYPE, NAME, AT_DISPATCH_CASE_QINT_TYPES_AND(SCALARTYPE, __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_QINT_BYTE_TYPES(...)               \
+  AT_DISPATCH_CASE_QINT(at::kQInt8, at::qint8, __VA_ARGS__) \
+  AT_DISPATCH_CASE_QINT(at::kQUInt8, at::quint8, __VA_ARGS__)
+
+#define AT_DISPATCH_QINT_BYTE_TYPES(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(TYPE, NAME, AT_DISPATCH_CASE_QINT_BYTE_TYPES(__VA_ARGS__))
+
+#define AT_DISPATCH_CASE_QINT_AND_SUB_BYTE_TYPES(...)                     \
+  AT_QINT_SUB_BYTE_PRIVATE_CASE_TYPE(                                     \
+      at::kQInt8, at::qint8, CHAR_BIT, SCHAR_MIN, SCHAR_MAX, __VA_ARGS__) \
+  AT_QINT_SUB_BYTE_PRIVATE_CASE_TYPE(                                     \
+      at::kQUInt8, at::quint8, CHAR_BIT, 0, UCHAR_MAX, __VA_ARGS__)       \
+  AT_QINT_SUB_BYTE_PRIVATE_CASE_TYPE(                                     \
+      at::kQInt32,                                                        \
+      at::qint32,                                                         \
+      CHAR_BIT * sizeof(int),                                             \
+      INT_MIN,                                                            \
+      INT_MAX,                                                            \
+      __VA_ARGS__)                                                        \
+  AT_QINT_SUB_BYTE_PRIVATE_CASE_TYPE(                                     \
+      at::kQUInt4x2, at::quint4x2, 4, 0, 15, __VA_ARGS__)                 \
+  AT_QINT_SUB_BYTE_PRIVATE_CASE_TYPE(                                     \
+      at::kQUInt2x4, at::quint2x4, 2, 0, 3, __VA_ARGS__)
+
+#define AT_DISPATCH_QINT_AND_SUB_BYTE_TYPES(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                        \
+      TYPE, NAME, AT_DISPATCH_CASE_QINT_AND_SUB_BYTE_TYPES(__VA_ARGS__))
+
+#define AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX(...) \
+  AT_DISPATCH_CASE_ALL_TYPES(__VA_ARGS__)           \
+  AT_DISPATCH_CASE_COMPLEX_TYPES(__VA_ARGS__)
+
+#define AT_DISPATCH_ALL_TYPES_AND_COMPLEX(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                      \
+      TYPE, NAME, AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX(__VA_ARGS__))
+
+#define AT_DISPATCH_CASE_ALL_TYPES_AND(SCALARTYPE, ...) \
+  AT_DISPATCH_CASE_ALL_TYPES(__VA_ARGS__)               \
+  AT_DISPATCH_CASE(SCALARTYPE, __VA_ARGS__)
+
+#define AT_DISPATCH_ALL_TYPES_AND(SCALARTYPE, TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                          \
+      TYPE, NAME, AT_DISPATCH_CASE_ALL_TYPES_AND(SCALARTYPE, __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX_AND(SCALARTYPE, ...) \
+  AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX(__VA_ARGS__)               \
+  AT_DISPATCH_CASE(SCALARTYPE, __VA_ARGS__)
+
+#define AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND(SCALARTYPE, TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                                      \
+      TYPE,                                                                \
+      NAME,                                                                \
+      AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX_AND(SCALARTYPE, __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_ALL_TYPES_AND2(SCALARTYPE1, SCALARTYPE2, ...) \
+  AT_DISPATCH_CASE_ALL_TYPES(__VA_ARGS__)                              \
+  AT_DISPATCH_CASE(SCALARTYPE1, __VA_ARGS__)                           \
+  AT_DISPATCH_CASE(SCALARTYPE2, __VA_ARGS__)
+
+#define AT_DISPATCH_ALL_TYPES_AND2(SCALARTYPE1, SCALARTYPE2, TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                                         \
+      TYPE,                                                                   \
+      NAME,                                                                   \
+      AT_DISPATCH_CASE_ALL_TYPES_AND2(SCALARTYPE1, SCALARTYPE2, __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX_AND2(  \
+    SCALARTYPE1, SCALARTYPE2, ...)                    \
+  AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX(__VA_ARGS__) \
+  AT_DISPATCH_CASE(SCALARTYPE1, __VA_ARGS__)          \
+  AT_DISPATCH_CASE(SCALARTYPE2, __VA_ARGS__)
+
+#define AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND2(    \
+    SCALARTYPE1, SCALARTYPE2, TYPE, NAME, ...)     \
+  AT_DISPATCH_SWITCH(                              \
+      TYPE,                                        \
+      NAME,                                        \
+      AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX_AND2( \
+          SCALARTYPE1, SCALARTYPE2, __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_ALL_TYPES_AND3(        \
+    SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, ...) \
+  AT_DISPATCH_CASE_ALL_TYPES(__VA_ARGS__)       \
+  AT_DISPATCH_CASE(SCALARTYPE1, __VA_ARGS__)    \
+  AT_DISPATCH_CASE(SCALARTYPE2, __VA_ARGS__)    \
+  AT_DISPATCH_CASE(SCALARTYPE3, __VA_ARGS__)
+
+#define AT_DISPATCH_ALL_TYPES_AND3(                         \
+    SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                       \
+      TYPE,                                                 \
+      NAME,                                                 \
+      AT_DISPATCH_CASE_ALL_TYPES_AND3(                      \
+          SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX_AND3(  \
+    SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, ...)       \
+  AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX(__VA_ARGS__) \
+  AT_DISPATCH_CASE(SCALARTYPE1, __VA_ARGS__)          \
+  AT_DISPATCH_CASE(SCALARTYPE2, __VA_ARGS__)          \
+  AT_DISPATCH_CASE(SCALARTYPE3, __VA_ARGS__)
+
+#define AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND3(             \
+    SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                       \
+      TYPE,                                                 \
+      NAME,                                                 \
+      AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX_AND3(          \
+          SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX_AND4(         \
+    SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, SCALARTYPE4, ...) \
+  AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX(__VA_ARGS__)        \
+  AT_DISPATCH_CASE(SCALARTYPE1, __VA_ARGS__)                 \
+  AT_DISPATCH_CASE(SCALARTYPE2, __VA_ARGS__)                 \
+  AT_DISPATCH_CASE(SCALARTYPE3, __VA_ARGS__)                 \
+  AT_DISPATCH_CASE(SCALARTYPE4, __VA_ARGS__)
+
+#define AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND4(                          \
+    SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, SCALARTYPE4, TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                                    \
+      TYPE,                                                              \
+      NAME,                                                              \
+      AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX_AND4(                       \
+          SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, SCALARTYPE4, __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX_AND5(                      \
+    SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, SCALARTYPE4, SCALARTYPE5, ...) \
+  AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX(__VA_ARGS__)                     \
+  AT_DISPATCH_CASE(SCALARTYPE1, __VA_ARGS__)                              \
+  AT_DISPATCH_CASE(SCALARTYPE2, __VA_ARGS__)                              \
+  AT_DISPATCH_CASE(SCALARTYPE3, __VA_ARGS__)                              \
+  AT_DISPATCH_CASE(SCALARTYPE4, __VA_ARGS__)                              \
+  AT_DISPATCH_CASE(SCALARTYPE5, __VA_ARGS__)
+
+#define AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND5(    \
+    SCALARTYPE1,                                   \
+    SCALARTYPE2,                                   \
+    SCALARTYPE3,                                   \
+    SCALARTYPE4,                                   \
+    SCALARTYPE5,                                   \
+    TYPE,                                          \
+    NAME,                                          \
+    ...)                                           \
+  AT_DISPATCH_SWITCH(                              \
+      TYPE,                                        \
+      NAME,                                        \
+      AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX_AND5( \
+          SCALARTYPE1,                             \
+          SCALARTYPE2,                             \
+          SCALARTYPE3,                             \
+          SCALARTYPE4,                             \
+          SCALARTYPE5,                             \
+          __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX_AND6(  \
+    SCALARTYPE1,                                      \
+    SCALARTYPE2,                                      \
+    SCALARTYPE3,                                      \
+    SCALARTYPE4,                                      \
+    SCALARTYPE5,                                      \
+    SCALARTYPE6,                                      \
+    ...)                                              \
+  AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX(__VA_ARGS__) \
+  AT_DISPATCH_CASE(SCALARTYPE1, __VA_ARGS__)          \
+  AT_DISPATCH_CASE(SCALARTYPE2, __VA_ARGS__)          \
+  AT_DISPATCH_CASE(SCALARTYPE3, __VA_ARGS__)          \
+  AT_DISPATCH_CASE(SCALARTYPE4, __VA_ARGS__)          \
+  AT_DISPATCH_CASE(SCALARTYPE5, __VA_ARGS__)          \
+  AT_DISPATCH_CASE(SCALARTYPE6, __VA_ARGS__)
+
+#define AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND6(    \
+    SCALARTYPE1,                                   \
+    SCALARTYPE2,                                   \
+    SCALARTYPE3,                                   \
+    SCALARTYPE4,                                   \
+    SCALARTYPE5,                                   \
+    SCALARTYPE6,                                   \
+    TYPE,                                          \
+    NAME,                                          \
+    ...)                                           \
+  AT_DISPATCH_SWITCH(                              \
+      TYPE,                                        \
+      NAME,                                        \
+      AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX_AND6( \
+          SCALARTYPE1,                             \
+          SCALARTYPE2,                             \
+          SCALARTYPE3,                             \
+          SCALARTYPE4,                             \
+          SCALARTYPE5,                             \
+          SCALARTYPE6,                             \
+          __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX_AND7(  \
+    SCALARTYPE1,                                      \
+    SCALARTYPE2,                                      \
+    SCALARTYPE3,                                      \
+    SCALARTYPE4,                                      \
+    SCALARTYPE5,                                      \
+    SCALARTYPE6,                                      \
+    SCALARTYPE7,                                      \
+    ...)                                              \
+  AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX(__VA_ARGS__) \
+  AT_DISPATCH_CASE(SCALARTYPE1, __VA_ARGS__)          \
+  AT_DISPATCH_CASE(SCALARTYPE2, __VA_ARGS__)          \
+  AT_DISPATCH_CASE(SCALARTYPE3, __VA_ARGS__)          \
+  AT_DISPATCH_CASE(SCALARTYPE4, __VA_ARGS__)          \
+  AT_DISPATCH_CASE(SCALARTYPE5, __VA_ARGS__)          \
+  AT_DISPATCH_CASE(SCALARTYPE6, __VA_ARGS__)          \
+  AT_DISPATCH_CASE(SCALARTYPE7, __VA_ARGS__)
+
+#define AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND7(    \
+    SCALARTYPE1,                                   \
+    SCALARTYPE2,                                   \
+    SCALARTYPE3,                                   \
+    SCALARTYPE4,                                   \
+    SCALARTYPE5,                                   \
+    SCALARTYPE6,                                   \
+    SCALARTYPE7,                                   \
+    TYPE,                                          \
+    NAME,                                          \
+    ...)                                           \
+  AT_DISPATCH_SWITCH(                              \
+      TYPE,                                        \
+      NAME,                                        \
+      AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX_AND7( \
+          SCALARTYPE1,                             \
+          SCALARTYPE2,                             \
+          SCALARTYPE3,                             \
+          SCALARTYPE4,                             \
+          SCALARTYPE5,                             \
+          SCALARTYPE6,                             \
+          SCALARTYPE7,                             \
+          __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX_AND8(  \
+    SCALARTYPE1,                                      \
+    SCALARTYPE2,                                      \
+    SCALARTYPE3,                                      \
+    SCALARTYPE4,                                      \
+    SCALARTYPE5,                                      \
+    SCALARTYPE6,                                      \
+    SCALARTYPE7,                                      \
+    SCALARTYPE8,                                      \
+    ...)                                              \
+  AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX(__VA_ARGS__) \
+  AT_DISPATCH_CASE(SCALARTYPE1, __VA_ARGS__)          \
+  AT_DISPATCH_CASE(SCALARTYPE2, __VA_ARGS__)          \
+  AT_DISPATCH_CASE(SCALARTYPE3, __VA_ARGS__)          \
+  AT_DISPATCH_CASE(SCALARTYPE4, __VA_ARGS__)          \
+  AT_DISPATCH_CASE(SCALARTYPE5, __VA_ARGS__)          \
+  AT_DISPATCH_CASE(SCALARTYPE6, __VA_ARGS__)          \
+  AT_DISPATCH_CASE(SCALARTYPE7, __VA_ARGS__)          \
+  AT_DISPATCH_CASE(SCALARTYPE8, __VA_ARGS__)
+
+#define AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND8(    \
+    SCALARTYPE1,                                   \
+    SCALARTYPE2,                                   \
+    SCALARTYPE3,                                   \
+    SCALARTYPE4,                                   \
+    SCALARTYPE5,                                   \
+    SCALARTYPE6,                                   \
+    SCALARTYPE7,                                   \
+    SCALARTYPE8,                                   \
+    TYPE,                                          \
+    NAME,                                          \
+    ...)                                           \
+  AT_DISPATCH_SWITCH(                              \
+      TYPE,                                        \
+      NAME,                                        \
+      AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX_AND8( \
+          SCALARTYPE1,                             \
+          SCALARTYPE2,                             \
+          SCALARTYPE3,                             \
+          SCALARTYPE4,                             \
+          SCALARTYPE5,                             \
+          SCALARTYPE6,                             \
+          SCALARTYPE7,                             \
+          SCALARTYPE8,                             \
+          __VA_ARGS__))
+
+#define AT_DISPATCH_CASE_BIT_TYPES(...)                  \
+  AT_DISPATCH_CASE(at::ScalarType::Bits1x8, __VA_ARGS__) \
+  AT_DISPATCH_CASE(at::ScalarType::Bits2x4, __VA_ARGS__) \
+  AT_DISPATCH_CASE(at::ScalarType::Bits4x2, __VA_ARGS__) \
+  AT_DISPATCH_CASE(at::ScalarType::Bits8, __VA_ARGS__)   \
+  AT_DISPATCH_CASE(at::ScalarType::Bits16, __VA_ARGS__)
+
+#define AT_DISPATCH_BIT_TYPES(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(TYPE, NAME, AT_DISPATCH_CASE_BIT_TYPES(__VA_ARGS__))
+
+#define AT_DISPATCH_INDEX_TYPES(TYPE, NAME, ...)     \
+  AT_DISPATCH_SWITCH(                                \
+      TYPE,                                          \
+      NAME,                                          \
+      AT_PRIVATE_CASE_TYPE_USING_HINT(               \
+          at::ScalarType::Int, index_t, __VA_ARGS__) \
+          AT_PRIVATE_CASE_TYPE_USING_HINT(           \
+              at::ScalarType::Long, index_t, __VA_ARGS__))
+
+// ----------------------------------------------------------------------------
+// DEPRECATED MACROS, DON'T USE THESE
+// ----------------------------------------------------------------------------
+
+#define AT_DISPATCH_ALL_TYPES_AND_HALF(TYPE, NAME, ...) \
+  detail::deprecated_AT_DISPATCH_ALL_TYPES_AND_HALF();  \
+  AT_DISPATCH_SWITCH(                                   \
+      TYPE,                                             \
+      NAME,                                             \
+      AT_DISPATCH_CASE_ALL_TYPES_AND(at::ScalarType::Half, __VA_ARGS__))

.venv/lib/python3.11/site-packages/torch/include/ATen/EmptyTensor.h

@@ -0,0 +1,166 @@
+#pragma once
+#include <ATen/core/TensorBase.h>
+
+namespace at::detail {
+
+inline void check_size_nonnegative(ArrayRef<int64_t> size) {
+  for (const auto& x : size) {
+    TORCH_CHECK(
+        x >= 0,
+        "Trying to create tensor with negative dimension ",
+        x,
+        ": ",
+        size);
+  }
+}
+
+inline void check_size_nonnegative(ArrayRef<c10::SymInt> size) {
+  for (const auto& x : size) {
+    TORCH_CHECK(
+        x.expect_size(__FILE__, __LINE__),
+        "Trying to create tensor with negative dimension ",
+        x,
+        ": ",
+        size);
+  }
+}
+
+TORCH_API size_t computeStorageNbytesContiguous(
+    IntArrayRef sizes,
+    size_t itemsize,
+    size_t storage_offset = 0);
+TORCH_API SymInt computeStorageNbytesContiguous(
+    SymIntArrayRef sizes,
+    const SymInt& itemsize,
+    const SymInt& storage_offset = 0);
+TORCH_API size_t computeStorageNbytes(
+    IntArrayRef sizes,
+    IntArrayRef strides,
+    size_t itemsize,
+    size_t storage_offset = 0);
+TORCH_API SymInt computeStorageNbytes(
+    SymIntArrayRef sizes,
+    SymIntArrayRef strides,
+    const SymInt& itemsize,
+    const SymInt& storage_offset = 0);
+
+TORCH_API TensorBase empty_generic(
+    IntArrayRef size,
+    c10::Allocator* allocator,
+    c10::DispatchKeySet ks,
+    ScalarType scalar_type,
+    std::optional<c10::MemoryFormat> memory_format_opt);
+
+TORCH_API TensorBase empty_generic_symint(
+    SymIntArrayRef size,
+    c10::Allocator* allocator,
+    c10::DispatchKeySet ks,
+    ScalarType scalar_type,
+    std::optional<c10::MemoryFormat> memory_format_opt);
+
+TORCH_API TensorBase empty_strided_generic(
+    IntArrayRef size,
+    IntArrayRef stride,
+    c10::Allocator* allocator,
+    c10::DispatchKeySet ks,
+    ScalarType scalar_type);
+
+TORCH_API TensorBase empty_strided_symint_generic(
+    SymIntArrayRef size,
+    SymIntArrayRef stride,
+    c10::Allocator* allocator,
+    c10::DispatchKeySet ks,
+    ScalarType scalar_type);
+
+TORCH_API TensorBase empty_cpu(
+    IntArrayRef size,
+    ScalarType dtype,
+    bool pin_memory = false,
+    std::optional<c10::MemoryFormat> memory_format_opt = std::nullopt);
+
+TORCH_API TensorBase empty_cpu(
+    IntArrayRef size,
+    std::optional<ScalarType> dtype_opt,
+    std::optional<Layout> layout_opt,
+    std::optional<Device> device_opt,
+    std::optional<bool> pin_memory_opt,
+    std::optional<c10::MemoryFormat> memory_format_opt);
+
+TORCH_API TensorBase empty_cpu(IntArrayRef size, const TensorOptions& options);
+
+TORCH_API TensorBase empty_strided_cpu(
+    IntArrayRef size,
+    IntArrayRef stride,
+    ScalarType dtype,
+    bool pin_memory = false);
+
+TORCH_API TensorBase empty_strided_cpu(
+    IntArrayRef size,
+    IntArrayRef stride,
+    std::optional<ScalarType> dtype_opt,
+    std::optional<Layout> layout_opt,
+    std::optional<Device> device_opt,
+    std::optional<bool> pin_memory_opt);
+
+TORCH_API TensorBase empty_strided_cpu(
+    IntArrayRef size,
+    IntArrayRef stride,
+    const TensorOptions& options);
+
+TORCH_API TensorBase empty_meta(
+    IntArrayRef size,
+    ScalarType dtype,
+    std::optional<c10::MemoryFormat> memory_format_opt = std::nullopt);
+
+TORCH_API TensorBase empty_meta(
+    IntArrayRef size,
+    std::optional<ScalarType> dtype_opt,
+    std::optional<Layout> layout_opt,
+    std::optional<Device> device_opt,
+    std::optional<bool> pin_memory_opt,
+    std::optional<c10::MemoryFormat> memory_format_opt);
+
+TORCH_API TensorBase empty_symint_meta(
+    SymIntArrayRef size,
+    std::optional<ScalarType> dtype_opt,
+    std::optional<Layout> layout_opt,
+    std::optional<Device> device_opt,
+    std::optional<bool> pin_memory_opt,
+    std::optional<c10::MemoryFormat> memory_format_opt);
+
+TORCH_API TensorBase empty_meta(IntArrayRef size, const TensorOptions& options);
+
+TORCH_API TensorBase
+empty_strided_meta(IntArrayRef size, IntArrayRef stride, ScalarType dtype);
+
+TORCH_API TensorBase empty_strided_meta(
+    IntArrayRef size,
+    IntArrayRef stride,
+    std::optional<ScalarType> dtype_opt,
+    std::optional<Layout> layout_opt,
+    std::optional<Device> device_opt,
+    std::optional<bool> pin_memory_opt);
+
+TORCH_API TensorBase empty_strided_meta(
+    IntArrayRef size,
+    IntArrayRef stride,
+    const TensorOptions& options);
+
+TORCH_API TensorBase empty_strided_symint_meta(
+    SymIntArrayRef size,
+    SymIntArrayRef stride,
+    ScalarType dtype);
+
+TORCH_API TensorBase empty_strided_symint_meta(
+    SymIntArrayRef size,
+    SymIntArrayRef stride,
+    std::optional<ScalarType> dtype_opt,
+    std::optional<Layout> layout_opt,
+    std::optional<Device> device_opt);
+
+TORCH_API TensorBase empty_strided_symint_meta(
+    SymIntArrayRef size,
+    SymIntArrayRef stride,
+    const TensorOptions& options);
+
+} // namespace at::detail

.venv/lib/python3.11/site-packages/torch/include/ATen/ExpandBase.h

@@ -0,0 +1,30 @@
+#include <ATen/core/TensorBase.h>
+
+// Broadcasting utilities for working with TensorBase
+namespace at {
+namespace internal {
+TORCH_API TensorBase expand_slow_path(const TensorBase& self, IntArrayRef size);
+} // namespace internal
+
+inline c10::MaybeOwned<TensorBase> expand_size(
+    const TensorBase& self,
+    IntArrayRef size) {
+  if (size.equals(self.sizes())) {
+    return c10::MaybeOwned<TensorBase>::borrowed(self);
+  }
+  return c10::MaybeOwned<TensorBase>::owned(
+      at::internal::expand_slow_path(self, size));
+}
+c10::MaybeOwned<TensorBase> expand_size(TensorBase&& self, IntArrayRef size) =
+    delete;
+
+inline c10::MaybeOwned<TensorBase> expand_inplace(
+    const TensorBase& tensor,
+    const TensorBase& to_expand) {
+  return expand_size(to_expand, tensor.sizes());
+}
+c10::MaybeOwned<TensorBase> expand_inplace(
+    const TensorBase& tensor,
+    TensorBase&& to_expand) = delete;
+
+} // namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/FuncTorchTLS.h

@@ -0,0 +1,46 @@
+#pragma once
+
+#include <c10/macros/Macros.h>
+#include <memory>
+
+namespace at::functorch {
+
+// NOTE [functorch TLS in pytorch/pytorch]
+//
+// functorch lives out-of-tree. However, it has some TLS that needs to be
+// propagated. The solution for that is we store a pointer to the TLS
+// inside pytorch/pytorch and extend FuncTorchTLSBase inside functorch to
+// include whatever functorch needs.
+//
+// We need to store a pointer due to the indirection:
+// inside functorch, we will create a subclass of FunctorchTLSBase called
+// FuncTorchTLSImpl that actually contains metadata, like the DynamicLayerStack.
+// FuncTorchTLSBase doesn't have any metadata because it hasn't been defined
+// yet.
+//
+// Here in pytorch/pytorch, we will pass around FuncTorchTLSBase*, but inside
+// functorch, we will assign a FuncTorchTLSImpl* to the FunctorchTLSBase*.
+// We can't directly pass around FunctorchTLSBase (without a pointer) because
+// FuncTorchTLSImpl does not fit inside a FuncTorchTLSBase by virtue of having
+// more elements.
+struct TORCH_API FuncTorchTLSBase {
+  virtual ~FuncTorchTLSBase() = default;
+  virtual std::unique_ptr<FuncTorchTLSBase> deepcopy() const = 0;
+
+  virtual int64_t checkSupportsSingleLevelAutogradFunction() const = 0;
+  virtual void checkSupportsCppAutogradFunction() const = 0;
+  virtual void checkSupportsInplaceRequiresGrad() const = 0;
+  virtual void checkSupportsRetainGrad() const = 0;
+};
+
+// returns deepcopy of the functorch tls
+TORCH_API std::unique_ptr<FuncTorchTLSBase> getCopyOfFuncTorchTLS();
+
+// sets the functorch tls. always does a deep copy.
+TORCH_API void setFuncTorchTLS(
+    const std::shared_ptr<const FuncTorchTLSBase>& state);
+
+// get a mutable reference to the functorch tls
+TORCH_API std::unique_ptr<FuncTorchTLSBase>& functorchTLSAccessor();
+
+} // namespace at::functorch

.venv/lib/python3.11/site-packages/torch/include/ATen/FunctionalStorageImpl.h

@@ -0,0 +1,208 @@
+#pragma once
+
+#include <ATen/Tensor.h>
+
+#include <utility>
+
+namespace at::functionalization {
+
+// See Note [Functionalization Pass In Core]
+
+// ViewMeta is a class used by the functionalization pass to navigate between
+// a base tensor and a view tensor.
+// For example, if I call `b = a.view1(...)`
+// the functionalization pass will generate and store a ViewMeta on b that looks
+// like:
+//
+// ViewMeta(
+//   [<captures>](const Tensor& base, int64_t mutated_view_idx) {
+//     return base.view1(...);
+//   },
+//   [<captures>](const at::Tensor& base, const at::Tensor& mutated_view,
+//   int64_t mutated_view_idx) -> at::Tensor {
+//     return at::functionalization::impl::view1_inverse(base, mutated_view,
+//     ...);
+//   }
+//
+// The forward_fn lambda describes how to replay view1 on a tensor.
+//
+// The reverse_fn lambda describes how, given a tensor that is already a view,
+// how to get the corresponding base tensor. See Note [Functionalization Pass:
+// View Inverses] for details.
+struct ViewMeta {
+  ViewMeta(
+      std::function<Tensor(const Tensor&, int64_t)> forward,
+      std::function<Tensor(const Tensor&, const Tensor&, int64_t)> reverse,
+      bool has_symbolic_inputs,
+      bool is_multi_output = false,
+      bool is_as_strided = false,
+      int64_t out_idx = 0)
+      : forward_fn(std::move(forward)),
+        reverse_fn(std::move(reverse)),
+        out_index(out_idx),
+        is_multi_output(is_multi_output),
+        is_as_strided(is_as_strided),
+        has_symbolic_inputs(has_symbolic_inputs) {}
+
+  std::function<Tensor(const Tensor&, int64_t)> forward_fn;
+  std::function<Tensor(const Tensor&, const Tensor&, int64_t)> reverse_fn;
+  // See Note [out_idx in ViewMeta]
+  int64_t out_index;
+
+  // Tells us if this is a multi-output view
+  bool is_multi_output;
+
+  bool is_as_strided;
+
+  // Tells us if this view operation has any symbolic inputs
+  bool has_symbolic_inputs;
+
+  // Returns a copy of the current ViewMeta, if out_idx matches the current
+  // out_index. Otherwise, returns a new ViewMeta with the same forward/reverse
+  // functions, but a new out index.
+  ViewMeta to_out_idx(int64_t out_idx);
+};
+
+// FunctionalStorageImpl is a subclass of StorageImpl used by the
+// functionalization pass. It has no underlying data (similar to meta storage).
+// It also knows how to reflect mutations to tensors in the absence of a valid
+// data pointer.
+//
+// A storage represents the state shared by (potentially multiple) views of the
+// same tensor. For example, in the following code:
+//
+// b = a.view1(...)
+// c = b.view2(...)
+// b.add_(1)
+// --> storage.add_update(b, {view1_meta})
+//
+// The call to add_(1) will result in a call to alias.add_update(b,
+// {view1_meta}), queueing up the mutation from b onto the alias. Later, suppose
+// c is used in an expression (e.g. you try to print c, or pass it to an
+// operator). Doing so will involve "syncing" c. First we apply any pending
+// updates to the alias, and then we regenerate c by replaying its views off of
+// the updated alias. E.g:
+//
+// print(str(c))
+// --> c.sync_()
+//     --> alias.apply_updates() // after this, the alias will be updated to
+//     reflect the mutation to b
+struct TORCH_API FunctionalStorageImpl : public c10::StorageImpl {
+ public:
+  struct Update {
+    // NOLINTNEXTLINE(cppcoreguidelines-avoid-const-or-ref-data-members)
+    const at::Tensor new_val;
+    // NOLINTNEXTLINE(cppcoreguidelines-avoid-const-or-ref-data-members)
+    const std::vector<ViewMeta> view_metas;
+  };
+
+  explicit FunctionalStorageImpl(const Tensor& value);
+
+  void add_update(
+      const Tensor& updated_val,
+      const std::vector<ViewMeta>& view_metas);
+  bool apply_updates();
+  const Tensor& base() {
+    return base_;
+  }
+  size_t generation() const {
+    return generation_;
+  }
+  void freeze() {
+    frozen_ = true;
+  }
+
+  c10::SymInt get_storage_size(bool before) {
+    if (before) {
+      return original_storage_size_;
+    } else {
+      return curr_storage_size_;
+    }
+  }
+
+  ~FunctionalStorageImpl() override = default;
+
+  void mark_mutation() {
+    mutation_counter_++;
+  }
+  void mark_mutation_during_no_grad_or_inference_mode() {
+    mutation_counter_during_no_grad_or_inference_mode_++;
+  }
+  void mark_mutation_hidden_from_autograd() {
+    mutation_counter_hidden_from_autograd_++;
+  }
+
+  bool are_all_mutations_under_no_grad_or_inference_mode() const {
+    auto non_autograd_mutations =
+        mutation_counter_during_no_grad_or_inference_mode_ +
+        mutation_counter_hidden_from_autograd_;
+    // The <= is because both counters will technically be incremented, if we
+    // perform e.g. a triton kernel mutation under no_grad
+    return mutation_counter_ <= non_autograd_mutations;
+  }
+
+  bool are_all_mutations_hidden_from_autograd() const {
+    // mutations under no_grad / inference_mode are technically not hidden from
+    // autograd - they change the version counter
+    return mutation_counter_ <= mutation_counter_hidden_from_autograd_;
+  }
+
+  void mark_inductor_storage_resize(c10::SymInt new_size) {
+    inductor_storage_resized_ = true;
+    curr_storage_size_ = std::move(new_size);
+  }
+
+  bool was_inductor_storage_resized() {
+    return inductor_storage_resized_;
+  }
+
+ private:
+  // NB: base_ should always point to a tensor BELOW the current
+  // functionalization layer. This is mainly to avoid reference cycles. e.g.
+  // given `b = a.view(...)` Both a.storage_ and b.storage_ are a
+  // FunctionStorageImpl containing an Walualias, with contains a Tensor
+  // `base_`. In this case (where a and b are FunctionalTensorWrapper's), base_
+  // should point not to a, but to a's unwrapped value, a.value_` See Note
+  // [Functionalization: Walualias Removal] for a diagram that shows this
+  // visually.
+  at::Tensor base_;
+  std::vector<Update> updates_;
+  // generation_ gets incremented every time a mutation is queued onto the
+  // alias. It is used to determine if a given tensor is "up to date", or if it
+  // needs to be regenerated from the alias.
+  size_t generation_ = 0;
+  // If frozen, no more mutations are allowed on this storage.  Once frozen, a
+  // storage cannot be unfrozen.
+  bool frozen_ = false;
+
+  // These mutation counters are bumped on the storage
+  // whenever a FunctionalTensorWrapper experiences a mutation.
+  // When the mutation is under no_grad, or comes from a triton kernel, we also
+  // bump the corresponding during_no_grad or hidden_from_autograd counters. Why
+  // do we need to detect these two situations separately from "normal" input
+  // mutations? (1) "normal" input mutations can mutate autograd metadata like
+  // .grad_fn,
+  //     in which case they need to be replayed outside of the compiled graph
+  // (2) "no_grad" input mutations are generally safe to keep in the graph (and
+  // compile),
+  //     but they bump the tensor's VC, so we need to mark_dirty() on the inputs
+  //     in torch.compile
+  // (3) mutations that are fully hidden from autograd (e.g. from a triton
+  // kernel)
+  //     do not mutate any autograd state, and be fully kept in the graph
+  // When we detect that an input was mutated, we need to be able to tell if:
+  // (1) all of the mutations were from triton kernels
+  // (2) all of the mutations were under no_grad
+  uint64_t mutation_counter_during_no_grad_or_inference_mode_ = 0;
+  uint64_t mutation_counter_ = 0;
+  uint64_t mutation_counter_hidden_from_autograd_ = 0;
+
+  // Used to tell if:
+  // (1) There were any storage resizes on a graph input
+  // (2) The original/curr storage size tell us if these resizes result in a nop
+  bool inductor_storage_resized_ = false;
+  c10::SymInt original_storage_size_;
+  c10::SymInt curr_storage_size_;
+};
+
+} // namespace at::functionalization

.venv/lib/python3.11/site-packages/torch/include/ATen/FunctionalTensorWrapper.h

@@ -0,0 +1,454 @@
+
+#pragma once
+
+#include <ATen/ArrayRef.h>
+#include <ATen/FunctionalStorageImpl.h>
+#include <ATen/core/IListRef.h>
+#include <ATen/core/List.h>
+#include <ATen/core/boxing/BoxedKernel.h>
+#include <ATen/core/boxing/impl/boxing.h>
+#include <ATen/core/dispatch/Dispatcher.h>
+
+#include <c10/core/DispatchKey.h>
+
+namespace at {
+
+// Note [Functionalization Pass In Core]
+// The Functionalization pass is used to remove aliasing from a pytorch program.
+//
+// This is useful for backends that don't support aliasing, like XLA and Vulkan.
+// It's also necessary in order to remove mutation from a program, which is
+// needed in Functorch.
+//
+// Consider this program:
+// a = torch.ones(...)
+// b = a.view(...)
+// b.add_(1)
+//
+// In this program, b is meant to alias with a due to the use of view(). At the
+// end of the program, both a and b are full of 2's. However, backends that
+// don't support aliasing aren't able to correctly implement the view()
+// operator. Instead, they can opt into the Functionalization pass, which will
+// sit between the user and the backend, and provide the necessary aliasing
+// logic.
+//
+// The functionalization pass will turn the above program into a slightly
+// different program that has the same semantics, transparently to the user,
+// that backends like XLA/Vulkan are able to implement a = torch.ones(...) b =
+// a.view_copy(...)  # view() replaced with view_copy(). Backends like
+// XLA/Vulkan can implement this! b.add_(1) a.add_(1)  # Our functionalization
+// pass machinery knows that a and b are aliased - it applies b's mutation to a
+// too.
+//
+// So, how does the functionalization pass keep track of which tensors are
+// aliased? The pass works by wrapping EVERY tensor in the program inside of a
+// FunctionalTensorWrapper, which knows about its alias'd tensors.
+//
+// See Note [Functionalization: Alias Removal] for details on the aliasing
+// machinery. See Note [Functionalization: Mutation Removal] for details on
+// mutation removal.
+struct TORCH_API FunctionalTensorWrapper : public c10::TensorImpl {
+  explicit FunctionalTensorWrapper(const Tensor& value);
+  // Additional constructor to create a FunctionalTensorWrapper directly from an
+  // underlying tensor that was created from a view. For example, the code b =
+  // a.view1() will generate a constructor call to FunctionalTensorWrapper(b, a,
+  // view1_meta)
+  explicit FunctionalTensorWrapper(
+      const Tensor& view_value,
+      const FunctionalTensorWrapper* base,
+      const functionalization::ViewMeta& meta);
+
+  // Get the underlying, actual tensor, that doesn't know anything about
+  // functionalization.
+  const Tensor& value() const {
+    return value_;
+  };
+  // The concept of "level" is only ever important to functorch; it's exposed
+  // here as more of a hook for functorch to use.
+  int64_t level() const {
+    return level_;
+  };
+  void set_level(int64_t level) {
+    level_ = level;
+  }
+  bool has_metadata_mutation() const {
+    return has_metadata_mutation_;
+  };
+
+  void mark_mutation() {
+    functional_storage_impl()->mark_mutation();
+  }
+  // Denotes a mutation that's hidden from autograd,
+  // e.g. for the purposes of passing a tensor to a triton kernel
+  void mark_mutation_hidden_from_autograd() {
+    functional_storage_impl()->mark_mutation_hidden_from_autograd();
+  }
+  void mark_mutation_during_no_grad_or_inference_mode() {
+    functional_storage_impl()->mark_mutation_during_no_grad_or_inference_mode();
+  }
+  // Are all the mutations happening to the tensor hidden from autograd
+  bool are_all_mutations_hidden_from_autograd() const {
+    return functional_storage_impl()->are_all_mutations_hidden_from_autograd();
+  }
+  // Did all mutations happen under no_grad or inference_mode
+  // (We also need to ignore mutations fully hidden from autograd here)
+  bool are_all_mutations_under_no_grad_or_inference_mode() const {
+    return functional_storage_impl()
+        ->are_all_mutations_under_no_grad_or_inference_mode();
+  }
+
+  void maybe_mark_symbolic(const functionalization::ViewMeta& meta) {
+    is_symbolic_ = is_symbolic_ | meta.has_symbolic_inputs;
+  }
+
+  bool is_symbolic() const {
+    return is_symbolic_;
+  }
+
+  // Runs the forward_fn of every ViewMeta collected in the current instance
+  // to some other base.
+  Tensor apply_view_metas(const Tensor& base);
+
+  // Sync's the underlying tensor with its alias, if it's out of date. This
+  // involves two steps: 1) Apply any pending updates/mutations to the alias 2)
+  // Replay the views (if any) to regenerate the current tensor off of the
+  // updated alias.
+  void sync_();
+  // Performs step (1) of the sync. This is its own public API because it's
+  // needed by view_inplace ops like transpose_. See Note [Functionalization
+  // Pass - Inplace View Ops]
+  void regenerate_from_base();
+  // Performs step (2) of the sync. This is its own public API because it's
+  // needed by functorch. functorch wants to make sure that all input tensors to
+  // a functionalized program have been properly synced so it can properly
+  // propagate mutations to inputs. It can't just call sync_(), because the
+  // FunctionalTensorWrapper will look like it has no aliases and sync_ will be
+  // a noop. We use the reference count on storage_ to determine if the wrapper
+  // is aliased, and by the time functorch is ready to propagate updates to
+  // inputs, any intermediate views of the input created by the program will
+  // have been deallocated. This function also returns whether or not the base
+  // actually had any updates to apply.
+  bool apply_updates();
+  // Takes the current state of value_ and snapshots it, sending it as a pending
+  // update to the alias.
+  void commit_update();
+  // When any tensor is mutated, the tensor increments its alias's "generation".
+  // Separately, each tensor maintains its own "generation" counter, which is
+  // used to determine if it's up-to-date with its alias. The act of syncing a
+  // tensor will set a tensor's generation equal to its alias's generation.
+  bool is_up_to_date() const;
+  // Freezes the storage of this tensor, preventing subsequent mutations
+  void freeze_storage() const;
+  // Every FunctionalTensorWrapper contains a vector<ViewMeta> objects
+  // describing the series of view ops that ran to generate the current tensor
+  // from the base tensor. This method is used by inplace-view ops like
+  // transpose_. It appends a ViewMeta to the existing stack, and refreshes the
+  // tensor by replaying the views off of the alias.
+  void mutate_view_meta(const at::functionalization::ViewMeta& meta);
+
+  // Custom implementation of self.set_(src)
+  void set__impl(const FunctionalTensorWrapper* other);
+
+  // Custom implementation of resize_storage_bytes_(self, new_size)
+  void storage_resize_(const c10::SymInt& new_size);
+
+  // Returns whether the current tensor's data was ever mutated
+  bool has_data_mutation();
+  //
+  // Returns whether the current FunctionalTensorWrapper
+  // experienced a set_() call.
+  bool was_storage_changed() {
+    return was_storage_changed_;
+  }
+
+  void set_storage_changed() {
+    was_storage_changed_ = true;
+  }
+
+  // A FunctionalTensor is considered a base if its not a view of another
+  // tensor.
+  bool isBaseTensor() const {
+    return view_metas_.empty();
+  }
+
+  c10::SymInt get_storage_size(bool before) {
+    return functional_storage_impl()->get_storage_size(before);
+  }
+
+  // Returns whether the FunctionalTensor experienced an
+  // untyped_storage().resize_() call
+  bool was_inductor_storage_resized() {
+    return functional_storage_impl()->was_inductor_storage_resized();
+  }
+
+  // The functionalization pass can be used to remove mutations.
+  // It does so by replacing any mutation op with it's corresponding
+  // out-of-place op, followed by a call to replace_(). e.g:
+  //
+  // a.add_(1)
+  //
+  // will turn into:
+  //
+  // tmp = a.add(1)
+  // a.replace_(tmp)
+  //
+  // replace_() swaps out the wrapped tensor, value_, with tmp.
+  void replace_(const Tensor& other, bool from_lazy_regenerate = false);
+
+  bool is_multi_output_view() {
+    return is_multi_output_view_;
+  }
+
+  // See Note[resize_() in functionalization pass]
+  void maybe_replace_storage(const Tensor& other);
+
+  // Replaces the storage with a new functional storage,
+  // and clears the view_metas_ stack.
+  // WARNING: Calling this function will sever the aliasing relationship between
+  // the current FunctionalTensorWrapper and any of its outstanding aliases.
+  // Please only call if you know what you're doing.
+  void _unsafe_reset_storage();
+
+  c10::intrusive_ptr<TensorImpl> shallow_copy_and_detach(
+      const c10::VariableVersion& version_counter,
+      bool allow_tensor_metadata_change) const override;
+
+  c10::intrusive_ptr<TensorImpl> shallow_copy_and_detach(
+      c10::VariableVersion&& version_counter,
+      bool allow_tensor_metadata_change) const override;
+
+  ~FunctionalTensorWrapper() override = default;
+
+  // FunctionalTensorWrapper overrides all custom size/stride function,
+  // so that if the inner tensor has a custom implementation
+  // we make sure to call that implementation.
+  at::IntArrayRef sizes_custom() const override;
+  at::IntArrayRef strides_custom() const override;
+  int64_t dim_custom() const override;
+  int64_t numel_custom() const override;
+  bool is_contiguous_custom(at::MemoryFormat memory_format) const override;
+  c10::SymIntArrayRef sym_sizes_custom() const override;
+  c10::SymInt sym_size_custom(int64_t d) const override;
+  c10::SymIntArrayRef sym_strides_custom() const override;
+  c10::SymInt sym_storage_offset_custom() const override;
+  c10::Device device_custom() const override;
+  c10::Layout layout_impl() const override;
+
+ private:
+  const char* tensorimpl_type_name() const override;
+  void set_constructor_metadata();
+  functionalization::FunctionalStorageImpl* functional_storage_impl() const;
+
+  // This is used to re-implement shallow_copy_and_detach for
+  // FunctionalTensorWrapper. The implementation is identical, but we just need
+  // to return a subclass instead of a plain TensorImpl.
+  // TODO: maybe it's possible to arrange for that to happen automatically
+  // without an override here?
+  template <typename VariableVersion>
+  c10::intrusive_ptr<TensorImpl> shallow_copy_and_detach_core(
+      VariableVersion&& version_counter,
+      bool allow_tensor_metadata_change) const;
+
+  void shallow_copy_from(const c10::intrusive_ptr<TensorImpl>& impl) override;
+  void copy_tensor_metadata_and_refresh(
+      const FunctionalTensorWrapper* src_impl,
+      FunctionalTensorWrapper* dest_impl,
+      const c10::VariableVersion& version_counter,
+      bool allow_tensor_metadata_change) const;
+
+  // Note that value is not taken by reference: internally, the wrapper will
+  // change the value tensor that it points to over time.
+  Tensor value_;
+  int64_t level_{};
+  // These two counters are used for identifying
+  // whether all the mutations on a given tensor are hidden from autograd or
+  // not. If we have an input mutation that is hidden from autograd, then once
+  // we convert the input mutation to a copy_() we know it will be safe to hide
+  // the copy_() from autograd as well.
+  bool has_metadata_mutation_ = false;
+  bool is_multi_output_view_ = false;
+  // Did the tensor experience a set_() call.
+  bool was_storage_changed_ = false;
+  // Did the tensor experience any view operation with symbolic int.
+  bool is_symbolic_ = false;
+
+  size_t generation_ = 0;
+  std::vector<at::functionalization::ViewMeta> view_metas_;
+
+ protected:
+  static void copy_tensor_metadata(
+      const FunctionalTensorWrapper* src_impl,
+      FunctionalTensorWrapper* dest_impl,
+      const c10::VariableVersion& version_counter,
+      bool allow_tensor_metadata_change);
+};
+
+// Utility functions for the functionalization pass.
+
+namespace functionalization {
+namespace impl {
+
+TORCH_API inline FunctionalTensorWrapper* unsafeGetFunctionalWrapper(
+    const Tensor& tensor) {
+  auto functional_impl =
+      static_cast<FunctionalTensorWrapper*>(tensor.unsafeGetTensorImpl());
+  TORCH_INTERNAL_ASSERT_DEBUG_ONLY(functional_impl != nullptr);
+  return functional_impl;
+}
+
+TORCH_API bool isBaseTensor(const at::Tensor& tensor);
+
+TORCH_API bool isFunctionalTensor(const at::Tensor& tensor);
+TORCH_API bool isFunctionalTensor(const std::optional<Tensor>& t);
+TORCH_API bool isFunctionalTensor(
+    const c10::List<std::optional<Tensor>>& t_list);
+TORCH_API bool isFunctionalTensor(ITensorListRef list);
+
+TORCH_API Tensor to_functional_tensor(const Tensor& tensor);
+TORCH_API std::optional<Tensor> to_functional_tensor(
+    const std::optional<Tensor>& tensor);
+TORCH_API c10::List<std::optional<Tensor>> to_functional_tensor(
+    const c10::List<std::optional<Tensor>>& t_list);
+TORCH_API std::vector<Tensor> to_functional_tensor(ITensorListRef t_list);
+
+TORCH_API void freeze_functional_tensor(const Tensor& tensor);
+
+TORCH_API Tensor
+from_functional_tensor(const Tensor& tensor, bool assert_functional = true);
+TORCH_API std::optional<Tensor> from_functional_tensor(
+    const std::optional<Tensor>& t,
+    bool assert_functional = true);
+TORCH_API c10::List<std::optional<Tensor>> from_functional_tensor(
+    const c10::List<std::optional<Tensor>>& t_list);
+TORCH_API std::vector<Tensor> from_functional_tensor(ITensorListRef t_list);
+
+TORCH_API void sync(const at::Tensor& t);
+TORCH_API void sync(const std::optional<Tensor>& t);
+TORCH_API void sync(const c10::List<std::optional<Tensor>>& t_list);
+TORCH_API void sync(ITensorListRef t_list);
+
+TORCH_API void replace_(const Tensor& functional_tensor, const Tensor& other);
+TORCH_API void replace_(
+    const ITensorListRef functional_tensor,
+    ITensorListRef other);
+
+TORCH_API void commit_update(const Tensor& functional_tensor);
+TORCH_API void commit_update(ITensorListRef functional_tensor);
+
+TORCH_API void unsafe_reset_storage(const Tensor& functional_tensor);
+
+TORCH_API void mark_mutation_hidden_from_autograd(
+    const Tensor& functional_tensor);
+
+TORCH_API bool are_all_mutations_hidden_from_autograd(
+    const Tensor& functional_tensor);
+
+TORCH_API bool are_all_mutations_under_no_grad_or_inference_mode(
+    const Tensor& functional_tensor);
+
+// These two methods are XLA-specific logic and are no-ops
+// for the normal functionalization flow.
+TORCH_API void propagate_xla_data(
+    const Tensor& functional_tensor,
+    const Tensor& other);
+TORCH_API void propagate_xla_data(
+    const ITensorListRef functional_tensor,
+    ITensorListRef other);
+
+TORCH_API void propagate_xla_data_direct(
+    const Tensor& tensor,
+    const Tensor& other);
+TORCH_API void propagate_xla_data_direct(
+    const ITensorListRef tensor,
+    ITensorListRef other);
+
+Tensor create_functional_tensor_with_view_meta(
+    const Tensor& view_to_wrap,
+    const Tensor& base,
+    functionalization::ViewMeta meta,
+    int64_t out_idx = 0);
+std::vector<Tensor> create_functional_tensor_with_view_meta(
+    ITensorListRef view_to_wrap,
+    const Tensor& base,
+    const functionalization::ViewMeta& meta);
+
+void mutate_view_meta(
+    const Tensor& self,
+    const functionalization::ViewMeta& meta);
+
+void set_sizes_strides_offset(const Tensor& out, const Tensor& meta_out);
+void set_sizes_strides_offset(
+    const std::vector<Tensor>& outs,
+    const std::vector<Tensor>& meta_outs);
+
+//  ~~~~~ TLS used in functionalization ~~~~~
+
+TORCH_API bool getFunctionalizationReapplyViewsTLS();
+TORCH_API void setFunctionalizationReapplyViewsTLS(bool reapply_views);
+
+class TORCH_API FunctionalizationReapplyViewsGuard {
+ public:
+  FunctionalizationReapplyViewsGuard(bool reapply_views)
+      : prev_(getFunctionalizationReapplyViewsTLS()) {
+    setFunctionalizationReapplyViewsTLS(reapply_views);
+  }
+
+  ~FunctionalizationReapplyViewsGuard() {
+    setFunctionalizationReapplyViewsTLS(prev_);
+  }
+
+  FunctionalizationReapplyViewsGuard(
+      const FunctionalizationReapplyViewsGuard&) = delete;
+  FunctionalizationReapplyViewsGuard operator=(
+      const FunctionalizationReapplyViewsGuard&) = delete;
+  FunctionalizationReapplyViewsGuard(FunctionalizationReapplyViewsGuard&&) =
+      delete;
+  FunctionalizationReapplyViewsGuard operator=(
+      FunctionalizationReapplyViewsGuard&&) = delete;
+
+ private:
+  bool prev_;
+};
+
+} // namespace impl
+
+// Helper function to call an out-of-place composite aten kernel that may use
+// mutations / views internally, and functionalize them.
+TORCH_API void functionalize_op_helper(
+    const c10::OperatorHandle& op,
+    torch::jit::Stack* stack);
+
+template <class Op, bool symint, class ReturnType, class... ParameterTypes>
+struct _functionalize_aten_op final {};
+
+template <class Op, bool symint, class ReturnType, class... ParameterTypes>
+struct _functionalize_aten_op<Op, symint, ReturnType(ParameterTypes...)> final {
+  static ReturnType call(
+      typename c10::maybe_keep_symint<symint, ParameterTypes>::type... args) {
+    using FuncType = ReturnType(
+        typename c10::maybe_keep_symint<symint, ParameterTypes>::type...);
+    auto op = c10::Dispatcher::singleton()
+                  .findSchemaOrThrow(
+                      (const char*)Op::name, (const char*)Op::overload_name)
+                  .typed<FuncType>();
+
+    return c10::impl::BoxedKernelWrapper<FuncType>::call(
+        c10::BoxedKernel::makeFromFunction<functionalize_op_helper>(),
+        op,
+        // BoxedKernelWrapper knows to ignore this keyset argument,
+        // because functionalize_op_helper doesn't take in a DispatchKeySet
+        c10::DispatchKeySet(),
+        args...);
+  }
+};
+
+template <class Op>
+using functionalize_aten_op =
+    _functionalize_aten_op<Op, false, typename Op::schema>;
+
+template <class Op>
+using functionalize_aten_op_symint =
+    _functionalize_aten_op<Op, true, typename Op::schema>;
+
+} // namespace functionalization
+} // namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/InferSize.h

@@ -0,0 +1,88 @@
+#pragma once
+
+#include <ATen/DimVector.h>
+#include <c10/core/ScalarType.h>
+#include <c10/core/SymIntArrayRef.h>
+#include <c10/util/DimVector.h>
+#include <optional>
+#include <sstream>
+#include <vector>
+
+namespace at {
+
+// Infers the size of a dim with size -1, if it exists. Also checks that new
+// shape is compatible with the number of elements.
+//
+// templated to handle std::vector<int64_t> and DimVector use cases, see
+// below
+//
+template <typename InputArrayRef, typename NumelType, typename ResultVec>
+inline void infer_size_impl(
+    InputArrayRef shape,
+    NumelType numel,
+    ResultVec& res) {
+  NumelType newsize = 1;
+  // N.B. this is an index, not a sym dim!
+  std::optional<int64_t> infer_dim;
+  for (int64_t dim = 0, ndim = shape.size(); dim != ndim; dim++) {
+    if (shape[dim] == -1) {
+      if (infer_dim) {
+        throw std::runtime_error("only one dimension can be inferred");
+      }
+      infer_dim = dim;
+    } else if (shape[dim] >= 0) {
+      newsize *= shape[dim];
+    } else {
+      AT_ERROR("invalid shape dimension ", shape[dim]);
+    }
+  }
+
+  if (TORCH_GUARD_SIZE_OBLIVIOUS(sym_eq(numel, newsize)) ||
+      (infer_dim && newsize > 0 && numel % newsize == 0)) {
+    if (infer_dim) {
+      // We have a degree of freedom here to select the dimension size; follow
+      // NumPy semantics and just bail.  However, a nice error message is needed
+      // because users often use `view` as a way to flatten & unflatten
+      // dimensions and will otherwise be confused why
+      //   empty_tensor.view( 0, 0)
+      // works yet
+      //   empty_tensor.view(-1, 0)
+      // doesn't.
+      TORCH_CHECK(
+          newsize != 0,
+          "cannot reshape tensor of 0 elements into shape ",
+          shape,
+          " because the unspecified dimension size -1 can be any "
+          "value and is ambiguous");
+      res[*infer_dim] = numel / newsize;
+    }
+    return;
+  }
+
+  std::ostringstream ss;
+  ss << "shape '" << shape << "' is invalid for input of size " << numel;
+  throw std::runtime_error(ss.str());
+}
+
+inline std::vector<int64_t> infer_size(IntArrayRef shape, int64_t numel) {
+  auto res = shape.vec();
+  infer_size_impl(shape, numel, res);
+  return res;
+}
+
+inline at::DimVector infer_size_dv(IntArrayRef shape, int64_t numel) {
+  auto res = at::DimVector(shape);
+  infer_size_impl(shape, numel, res);
+  return res;
+}
+
+inline at::SymDimVector infer_size_dv(
+    c10::SymIntArrayRef shape,
+    c10::SymInt numel) {
+  auto res = at::SymDimVector(shape);
+  infer_size_impl<c10::SymIntArrayRef, c10::SymInt, at::SymDimVector>(
+      shape, std::move(numel), res);
+  return res;
+}
+
+} // namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/InitialTensorOptions.h

@@ -0,0 +1,15 @@
+#pragma once
+
+#include <c10/core/TensorOptions.h>
+
+namespace at {
+
+// Represents the initial TensorOptions, before the "defaults" are ever changed.
+// This is designed to be used in library code, where the explicit devices,
+// dtypes, etc. are known. NOTE: this is not a stable API.
+inline TensorOptions initialTensorOptions() {
+  return TensorOptions(kCPU).dtype(kFloat).layout(kStrided).requires_grad(
+      false);
+}
+
+} // namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/Layout.h

@@ -0,0 +1,2 @@
+#pragma once
+#include <c10/core/Layout.h>

.venv/lib/python3.11/site-packages/torch/include/ATen/LegacyBatchedFallback.h

@@ -0,0 +1,25 @@
+#pragma once
+#include <ATen/ATen.h>
+#include <ATen/core/op_registration/op_registration.h>
+#include <torch/library.h>
+
+namespace at {
+
+// If an operator doesn't have a batching rule implemented then we fallback
+// to this implementation. The fallback only works on out-of-place operators
+// that return only tensors with new memory. (e.g., no in-place operators, no
+// view operations).
+//
+// The fallback effectively takes all of the BatchedTensors in `stack`, slices
+// them, and runs `op` on all of the corresponding slices to produce slices
+// of the outputs. The output slices then get `torch.stack`ed to create the
+// final returns.
+//
+// The performance of the fallback is not very good because it introduces an
+// extra copy from stacking the sliced outputs. Because of this, we prefer to
+// write batching rules for operators whenever possible.
+void batchedTensorForLoopFallback(
+    const c10::OperatorHandle& op,
+    torch::jit::Stack* stack);
+
+} // namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/LegacyBatchedTensorImpl.h

@@ -0,0 +1,160 @@
+#pragma once
+
+#include <bitset>
+
+#include <ATen/ArrayRef.h>
+#include <ATen/SmallVector.h>
+#include <ATen/Tensor.h>
+
+namespace at {
+
+// We assume this in a few other places in the codebase,
+// but there isn't a centralized definition.
+constexpr int64_t kVmapMaxTensorDims = 64;
+
+// The valid vmap levels range from [0, 64). This effectively means that we
+// support a maximum of 64 nested vmaps.
+constexpr int64_t kVmapNumLevels = 64;
+
+// Store this number of elements of BatchDims on the stack. Most people will
+// probably use <= 5 nested vmaps, but adjust this number as necessary.
+constexpr int64_t kBatchDimsStackSize = 5;
+
+// a BatchDim represents a "private" dimension on a Tensor created inside of
+// vmap. It is a (level, dim) tuple, with the `dim` indicating which dimension
+// is being vmap'ed over and the `level` being an identifier for which vmap
+// said dimension was created inside. The `dim` corresponds to a "physical
+// dim" - it is a dimension index on the underlying physical tensor that is
+// being vmapped over.
+struct BatchDim {
+  BatchDim(int64_t level, int64_t dim) : dim_(dim), level_(level) {}
+  int64_t dim() const {
+    return dim_;
+  }
+  int64_t level() const {
+    return level_;
+  }
+
+ private:
+  int64_t dim_;
+  int64_t level_;
+};
+
+using BatchDims = SmallVector<BatchDim, kBatchDimsStackSize>;
+using BatchDimsRef = ArrayRef<BatchDim>;
+
+// A BatchedTensorImpl holds an underlying Tensor and a list of BatchDim
+// NB: We use the term "BatchedTensor" to mean a Tensor that is backed with a
+// BatchedTensorImpl.
+//
+// The batch dimensions are treated as being "private"; they are not
+// user-visible. For example, in the following Tensor,
+//    bt = BatchedTensorImpl(ones(2, 3, 5, 7), [(lvl=1, dim=0), (lvl=2, dim=1)])
+// dimensions 0 and 1 are batch dimensions.
+//
+// bt.sizes() returns (5, 7); bt.sum(0) performs a reduction over the (public)
+// dim 0, which is equivalent to dim 3 in the underlying ones(2, 3, 5, 7)
+// tensor.
+struct TORCH_API BatchedTensorImpl : public c10::TensorImpl {
+  explicit BatchedTensorImpl(Tensor value, BatchDims bdims);
+
+  // Returns a reference to BatchDims that represent which dimensions of this
+  // tensor are private.
+  BatchDimsRef bdims() const {
+    return bdims_;
+  }
+
+  // BatchedTensorImpl wraps a Tensor
+  const Tensor& value() const {
+    return value_;
+  };
+
+  // Given a public dimension index, return the dimension index in the
+  // underlying value() tensor. For example, if we have
+  //    bt = BatchedTensorImpl(ones(2, 3, 5, 7), [(lvl=1, dim=0), (lvl=2,
+  //    dim=2)])
+  // bt.actualDim(0) -> 1
+  // bt.actualDim(1) -> 3
+  // bt.actualDim(2) -> Error
+  int64_t actualDim(int64_t dim, bool wrap_dim = true) const;
+
+  // We have to override this because we opted into CustomStrides
+  IntArrayRef strides_custom() const override;
+  // Override a bunch of methods inherited from TensorImpl to return error
+  // messages.
+  bool is_contiguous_custom(at::MemoryFormat memory_format) const override;
+  void set_size(int64_t dim, int64_t new_size) override;
+  void set_stride(int64_t dim, int64_t new_stride) override;
+  void set_storage_offset(int64_t storage_offset) override;
+#ifdef DEBUG
+  bool has_storage() const override;
+#endif
+
+ private:
+  // see NOTE: [BatchedTensorImpl levels invariant]
+  void checkInvariants() const;
+  const char* tensorimpl_type_name() const override;
+
+  Tensor value_;
+
+  // Note: [BatchedTensorImpl levels invariant]
+  // There is an invariant that the BatchDims must be stored in increasing
+  // `level` order. That is, for i < j, bdims_[i].level must be less than
+  // bdims_[j].level.
+  BatchDims bdims_;
+};
+
+// NB: We use the term "BatchedTensor" to mean a Tensor that is backed with a
+// BatchedTensorImpl.
+inline bool isBatchedTensor(const Tensor& tensor) {
+  return tensor.unsafeGetTensorImpl()->key_set().has(DispatchKey::Batched);
+}
+
+// It is unsafe to call this on a Tensor that is not backed by a
+// BatchedTensorImpl. Please use `maybeGetBatchedImpl` whenever possible.
+inline BatchedTensorImpl* unsafeGetBatchedImpl(const Tensor& tensor) {
+  return static_cast<BatchedTensorImpl*>(tensor.unsafeGetTensorImpl());
+}
+
+inline BatchedTensorImpl* maybeGetBatchedImpl(const Tensor& tensor) {
+  if (!isBatchedTensor(tensor)) {
+    return nullptr;
+  }
+  return unsafeGetBatchedImpl(tensor);
+}
+
+// Returns a bitset. If bit i is set, then that means dim i is a batchdim.
+inline std::bitset<kVmapMaxTensorDims> createBatchDimBitset(
+    BatchDimsRef bdims) {
+  std::bitset<kVmapMaxTensorDims> is_bdim;
+  for (const auto& bdim : bdims) {
+    is_bdim.set(bdim.dim());
+  }
+  return is_bdim;
+}
+
+// Creates a bitset for all of the levels present in `bdims`
+inline std::bitset<kVmapNumLevels> createVmapLevelsBitset(BatchDimsRef bdims) {
+  std::bitset<kVmapNumLevels> result;
+  for (const auto& bdim : bdims) {
+    result.set(bdim.level());
+  }
+  return result;
+}
+
+inline std::ostream& operator<<(std::ostream& out, const BatchDim& bdim) {
+  out << "(lvl=" << bdim.level() << ", dim=" << bdim.dim() << ")";
+  return out;
+}
+
+// Use this to construct a BatchedTensor from a regular Tensor
+TORCH_API Tensor makeBatched(const Tensor& tensor, BatchDims bdims);
+
+// Adds a batch dim to `tensor`, returning a BatchedTensor
+TORCH_API Tensor addBatchDim(const Tensor& tensor, int64_t level, int64_t dim);
+
+// Checks if an inplace operation on self and other is "vmap compatible".
+// See NOTE: [vmap-incompatible in-place operations] for the definition of this.
+TORCH_API bool inplaceIsVmapCompatible(const Tensor& self, const Tensor& other);
+
+} // namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/LegacyVmapMode.h

@@ -0,0 +1,26 @@
+#pragma once
+
+#include <c10/core/impl/LocalDispatchKeySet.h>
+
+namespace at::impl {
+
+// VmapMode contains a thread local count of how many nested vmaps
+// we are currently inside. That number is known as the `vmap level`.
+// VmapMode is used in the implementation of the Python `torch.vmap` API.
+//
+// NOTE: this is NOT the c++ api for torch.vmap. That doesn't exist yet.
+
+struct TORCH_API VmapMode {
+  // Returns the vmap level, aka the count of how many nested vmaps we're in.
+  static int64_t current_vmap_level();
+
+  // Increment the count of nested vmaps. If this causes the vmap level to be
+  // greater than 0, then it enables DispatchKey::VmapMode on all tensors.
+  static int64_t increment_nesting();
+
+  // Decrements the count of nested vmaps. If this causes the vmap level to be
+  // equal to 0, then it disables DispatchKey::VmapMode on all tensors.
+  static int64_t decrement_nesting();
+};
+
+} // namespace at::impl

.venv/lib/python3.11/site-packages/torch/include/ATen/LegacyVmapTransforms.h

@@ -0,0 +1,183 @@
+#pragma once
+
+#include <ATen/LegacyBatchedTensorImpl.h>
+#include <ATen/core/IListRef.h>
+
+namespace at {
+
+// This file contains abstractions used for transforming *logical* vmap
+// arguments into *physical* arguments. (Keep reading for definitions of these
+// terms).
+
+// NOTE: [Logical vs physical args]
+// Consider the following vmap.
+//   vmap(vmap(func, in_dims=(2,)), in_dims=(0,))(torch.ones(2, 3, 4))
+// This would produce a BatchedTensor wrapping a Tensor of size [2, 3, 4],
+// with batch dims 0 and 2:
+//   BatchedTensor(ones(2, 3, 4), bdims=[(lvl=1,dim=0),(lvl=2,dim=2)])
+//
+// We say the *logical* view of the tensor has size [3] -- tensors inside
+// `func` appear to have size [3].
+// However, the *physical* underlying tensor (the one passed to vmap) has size
+// [2, 3, 4].
+//
+// This notion of logical vs physical also extends to non-tensor arguments.
+// Consider the previous tensor; let's assume the user called
+// `torch.sum(tensor, dim=0)` inside of `func`. Then the logical
+// dimension they are reducing over is dim 0 but the physical dim is dim 1
+// (the first non-batch dimension)
+
+// Forward declared; see NOTE: [What is a VmapPhysicalView?]
+struct VmapPhysicalView;
+
+// Most PyTorch operators take 4 or fewer inputs.
+constexpr int64_t kVmapTransformStaticInputSize = 4;
+using VmapPhysicalViewVec =
+    SmallVector<VmapPhysicalView, kVmapTransformStaticInputSize>;
+
+// Pytorch generally advertises good performance for <= 5 dims.
+// (see ATen/core/DimVector.h). We add a few extra dims (~3) for vmap
+// dimensions to get 8. Adjust this number as necessary
+constexpr int64_t kVmapStaticDimVecSize = 8;
+using VmapDimVector = SmallVector<int64_t, kVmapStaticDimVecSize>;
+using VmapSymDimVector = SmallVector<c10::SymInt, kVmapStaticDimVecSize>;
+
+// NOTE: [What is an VmapTransform?]
+// An *VmapTransform* converts logical views of tensors to physical views.
+//
+// Batching rules use VmapTransforms to convert logical arguments to
+// physical arguments, then call one or more at:: operator that handles the
+// physical arguments, and then converts the physical result back to a logical
+// argument.
+
+// VmapTransform for operators that take tensors with multiple batch dims.
+// Given one or more logical views on Tensors, `logicalToPhysical`
+// permutes all of the batch dims to the front of the tensor, aligns
+// and expands the batch dims to match each other (according to their `level`),
+// and returns a VmapPhysicalView on the tensor(s).
+struct TORCH_API MultiBatchVmapTransform {
+  static VmapPhysicalView logicalToPhysical(const Tensor& logical_tensor);
+  static VmapPhysicalViewVec logicalToPhysical(ITensorListRef logical_tensors);
+};
+
+// VmapTransform for operators that broadcast all inputs.
+// Given some logical views on Tensors, `logicalToPhysical`:
+// - permutes all of the batch dims to the front of the tensors
+// - aligns all the batch dims to the collective levels of all of the tensors.
+//   If a tensor does not have a batch dim for a vmap level, then it receives
+//   a size-one dimension for said level.
+// - aligns the non-batch dims to have the same dimensionality, adding extra
+//   size-1 dimensions in between the batch dimensions and the non-batch
+//   dimensions so that the batch dimensions are lined up from the right.
+//
+// For example: given inputs of size (B, 2) and (B, 3, 2) where B is the batch
+// dimension, BroadcastingVmapTransform returns VmapPhysicalViews that wrap
+// tensors of size (B, 1, 2) and (B, 3, 2).
+//
+// Given inputs of size (B, 2) and (2,), BroadcastingVmapTransform returns
+// VmapPhysicalViews wrapping tensors of size (B, 2) and (1, 2). We don't
+// actually *need* to return a tensor of size (1, 2) for the second tensor
+// because the broadcasting operation takes care of that for us, but we do
+// it anyways to keep things simple.
+struct TORCH_API BroadcastingVmapTransform {
+  static VmapPhysicalViewVec logicalToPhysical(TensorList logical_tensors);
+};
+
+// Forward declared, if you're reading this file head to toe, don't worry about
+// it yet.
+struct VmapPhysicalToLogicalMap;
+
+// NOTE: [What is a VmapPhysicalView?]
+// VmapPhysicalView represents a physical view on a Tensor.
+//
+// One can use it to further convert logical dimension indices, logical shapes,
+// and more to their physical variants, or convert a new (physical) tensor into
+// a logical BatchedTensor. (TODO(rzou): some of these are not yet implemented).
+//
+// VmapPhysicalView stores a physical tensor with all of its batch dimensions at
+// the front and some levels that correspond to said batch dimensions.
+//
+// The levels bitset specifies which vmap levels correspond to the batch
+// dimensions at the front of the tensor. In particular, the number of set bits
+// corresponds to the number of batch dimensions on `tensor` and the rightmost
+// bit of `levels` specifies the maximum number of nested vmaps we are in at
+// this point in time.
+// For example, given:
+//   physical_view = VmapPhysicalView(tensor=ones(2, 3, 4, 5, 6), levels={1, 3})
+//
+// Rightmost bit of `levels` is 3 indicating the number of nested vmaps less
+// than or equal to 3.
+//   bitset: 010100
+//              ^
+//              |
+//   levels: 012345
+struct TORCH_API VmapPhysicalView {
+  VmapPhysicalView(Tensor&& tensor, std::bitset<kVmapNumLevels> levels)
+      : levels_(levels), tensor_(std::move(tensor)) {
+    TORCH_INTERNAL_ASSERT(!isBatchedTensor(tensor_));
+  }
+
+  Tensor& tensor() {
+    return tensor_;
+  }
+  const Tensor& tensor() const {
+    return tensor_;
+  }
+
+  // Maps logical dim indices to physical dim indices. Also does dim wrapping.
+  //
+  // For example, given:
+  //   physical_view = VmapPhysicalView(tensor=ones(2, 3, 4, 5), levels={1, 3})
+  //
+  // Then physical_view.getPhysicalDims({0, 1}) returns {2, 3}.
+  // This is because the size of levels tell us that the first two dimensions
+  // of `tensor_` are batch dimensions, so a logical dim of `n` is actually
+  // a physical dim of `n + 2`.
+  VmapDimVector getPhysicalDims(OptionalIntArrayRef logical_dims) const;
+  int64_t getPhysicalDim(int64_t logical_dim) const;
+
+  // Returns a VmapPhysicalToLogicalMap object. This can be used for
+  // mapping a physical tensor to a new logical tensor (BatchedTensor)
+  VmapPhysicalToLogicalMap getPhysicalToLogicalMap() const;
+
+  // Maps a logical shape to a physical shape by pre-pending the batch
+  // sizes to the logical shape.
+  VmapDimVector getPhysicalShape(IntArrayRef logical_shape) const;
+
+  int64_t numBatchDims() const;
+
+ private:
+  int64_t numLogicalDims() const;
+
+  std::bitset<kVmapNumLevels> levels_;
+  Tensor tensor_;
+};
+
+// Convenience struct used for mapping a physical tensor (a non-BatchedTensor)
+// to a logical one (BatchedTensor). It holds some levels that are used to do
+// the mapping and assumes that the batch dimensions in the physical tensor all
+// occur at the front of the tensor.
+struct TORCH_API VmapPhysicalToLogicalMap {
+  VmapPhysicalToLogicalMap(std::bitset<kVmapNumLevels> levels)
+      : levels_(levels) {}
+
+  // Maps a physical tensor to a new logical tensor (BatchedTensor).
+  // Assumes that all of the "batch dimensions" are at the front
+  // of the physical tensor. For example, given:
+  // - x = rank-4 Tensor with size 2, 3, 5, 7
+  // - levels = (2, 4)
+  // Returns:
+  // - BatchedTensor(x, bdims=[(dim=0,lvl=2), (dim=1, lvl=4)])
+  Tensor apply(const Tensor& physical_tensor) const;
+
+  // Given a vector of physical tensors,
+  // 1. maps each tensor to a new logical tensor. Assumes that all of the
+  //    "batch dimensions" are at the front of the physical tensors.
+  // 2. stores the new logical tensors back into the passed-in vector. This is
+  //    to avoid additional dynamic allocations.
+  void applyInplace(std::vector<Tensor>& physical_tensors) const;
+
+  std::bitset<kVmapNumLevels> levels_;
+};
+
+} // namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/MapAllocator.h

@@ -0,0 +1,143 @@
+#pragma once
+
+#include <c10/core/Allocator.h>
+#include <c10/util/string_view.h>
+
+namespace at {
+
+enum MappedAllocatorModes {
+  ALLOCATOR_MAPPED_SHARED = 1,
+  ALLOCATOR_MAPPED_SHAREDMEM = 2,
+  ALLOCATOR_MAPPED_EXCLUSIVE = 4,
+  ALLOCATOR_MAPPED_NOCREATE = 8,
+  ALLOCATOR_MAPPED_KEEPFD = 16,
+  ALLOCATOR_MAPPED_FROMFD = 32,
+  ALLOCATOR_MAPPED_UNLINK = 64
+};
+
+// Sentinel value/type to help distinguish the file descriptor constructor from
+// the non-file descriptor constructor
+enum WithFd { WITH_FD };
+
+TORCH_API std::string NewProcessWideShmHandle();
+
+class TORCH_API MapAllocator {
+ public:
+  MapAllocator(c10::string_view filename, int flags, size_t size);
+  MapAllocator(
+      WithFd,
+      c10::string_view filename,
+      int fd,
+      int flags,
+      size_t size);
+  MapAllocator(const MapAllocator&) = delete;
+  MapAllocator& operator=(const MapAllocator&) = delete;
+  MapAllocator(MapAllocator&&) = delete;
+  MapAllocator& operator=(MapAllocator&&) = delete;
+
+  const char* filename() const {
+    return filename_.c_str();
+  }
+  int fd() const {
+#ifdef _WIN32
+    TORCH_CHECK(false, "MapAllocator::fd() is unsupported on Windows");
+#else
+    return fd_;
+#endif
+  }
+  ptrdiff_t size() const {
+    return size_;
+  }
+  // Return a pointer to the actual data for this allocator
+  // (in the case of the refcounted allocator, this is offset
+  // from the base pointer.)
+  virtual void* data() const {
+    return base_ptr_;
+  }
+
+  int flags() const {
+    return flags_;
+  }
+
+  static MapAllocator* fromDataPtr(const at::DataPtr&);
+  static at::DataPtr makeDataPtr(
+      c10::string_view filename,
+      int flags,
+      size_t size,
+      size_t* actual_size_out);
+  static at::DataPtr makeDataPtr(
+      WithFd,
+      const char* filename,
+      int fd,
+      int flags,
+      size_t size,
+      size_t* actual_size_out);
+
+  // Closes the data.  Helps us avoid destructor shenanigans
+  virtual void close();
+
+  // This is very dangerous.  You have to redefine this destructor for each
+  // subclass
+  virtual ~MapAllocator();
+
+ protected:
+  bool closed_ = false;
+  std::string filename_;
+  int flags_ = 0;
+  ptrdiff_t size_; /* mapped size */
+#ifdef _WIN32
+  void* handle_;
+  void* event_;
+  std::string eventname_;
+#else
+  int fd_ = -1;
+#endif
+  void* base_ptr_ = nullptr;
+};
+
+// Base-from-member idiom
+struct TORCH_API RefcountedMapAllocatorArgCheck {
+  RefcountedMapAllocatorArgCheck(int flags);
+};
+
+class TORCH_API RefcountedMapAllocator : private RefcountedMapAllocatorArgCheck,
+                                         public MapAllocator {
+ public:
+  RefcountedMapAllocator(const char* filename, int flags, size_t size);
+  RefcountedMapAllocator(
+      WithFd,
+      const char* filename,
+      int fd,
+      int flags,
+      size_t size);
+
+  static RefcountedMapAllocator* fromDataPtr(const at::DataPtr&);
+  static at::DataPtr makeDataPtr(
+      const char* filename,
+      int flags,
+      size_t size,
+      size_t* actual_size_out);
+  static at::DataPtr makeDataPtr(
+      WithFd,
+      const char* filename,
+      int fd,
+      int flags,
+      size_t size,
+      size_t* actual_size_out);
+
+  void* data() const override;
+
+  void incref();
+  int decref();
+  void close() override;
+
+  ~RefcountedMapAllocator() override {
+    RefcountedMapAllocator::close();
+  }
+
+ protected:
+  void checkFlags();
+  void initializeAlloc();
+};
+
+} // namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/MatrixRef.h

@@ -0,0 +1,107 @@
+#pragma once
+#include <ATen/Utils.h>
+#include <c10/util/ArrayRef.h>
+
+namespace at {
+/// MatrixRef - Like an ArrayRef, but with an extra recorded strides so that
+/// we can easily view it as a multidimensional array.
+///
+/// Like ArrayRef, this class does not own the underlying data, it is expected
+/// to be used in situations where the data resides in some other buffer.
+///
+/// This is intended to be trivially copyable, so it should be passed by
+/// value.
+///
+/// For now, 2D only (so the copies are actually cheap, without having
+/// to write a SmallVector class) and contiguous only (so we can
+/// return non-strided ArrayRef on index).
+///
+/// P.S. dimension 0 indexes rows, dimension 1 indexes columns
+template <typename T>
+class MatrixRef {
+ public:
+  typedef size_t size_type;
+
+ private:
+  /// Underlying ArrayRef
+  ArrayRef<T> arr;
+
+  /// Stride of dim 0 (outer dimension)
+  size_type stride0;
+
+  // Stride of dim 1 is assumed to be 1
+
+ public:
+  /// Construct an empty Matrixref.
+  /*implicit*/ MatrixRef() : arr(nullptr), stride0(0) {}
+
+  /// Construct an MatrixRef from an ArrayRef and outer stride.
+  /*implicit*/ MatrixRef(ArrayRef<T> arr, size_type stride0)
+      : arr(arr), stride0(stride0) {
+    TORCH_CHECK(
+        arr.size() % stride0 == 0,
+        "MatrixRef: ArrayRef size ",
+        arr.size(),
+        " not divisible by stride ",
+        stride0)
+  }
+
+  /// @}
+  /// @name Simple Operations
+  /// @{
+
+  /// empty - Check if the matrix is empty.
+  bool empty() const {
+    return arr.empty();
+  }
+
+  const T* data() const {
+    return arr.data();
+  }
+
+  /// size - Get size a dimension
+  size_t size(size_t dim) const {
+    if (dim == 0) {
+      return arr.size() / stride0;
+    } else if (dim == 1) {
+      return stride0;
+    } else {
+      TORCH_CHECK(
+          0, "MatrixRef: out of bounds dimension ", dim, "; expected 0 or 1");
+    }
+  }
+
+  size_t numel() const {
+    return arr.size();
+  }
+
+  /// equals - Check for element-wise equality.
+  bool equals(MatrixRef RHS) const {
+    return stride0 == RHS.stride0 && arr.equals(RHS.arr);
+  }
+
+  /// @}
+  /// @name Operator Overloads
+  /// @{
+  ArrayRef<T> operator[](size_t Index) const {
+    return arr.slice(Index * stride0, stride0);
+  }
+
+  /// Disallow accidental assignment from a temporary.
+  ///
+  /// The declaration here is extra complicated so that "arrayRef = {}"
+  /// continues to select the move assignment operator.
+  template <typename U>
+  std::enable_if_t<std::is_same_v<U, T>, MatrixRef<T>>& operator=(
+      U&& Temporary) = delete;
+
+  /// Disallow accidental assignment from a temporary.
+  ///
+  /// The declaration here is extra complicated so that "arrayRef = {}"
+  /// continues to select the move assignment operator.
+  template <typename U>
+  std::enable_if_t<std::is_same_v<U, T>, MatrixRef<T>>& operator=(
+      std::initializer_list<U>) = delete;
+};
+
+} // end namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/MetaFunctions.h

@@ -0,0 +1,29 @@
+#include <ATen/core/TensorBody.h>
+
+// TODO Undo all logic introduced for Note [Avoiding Include Cycles In Static Dispatch]
+// Code introduced to avoid cyclic dependency in static dispatch is no longer
+// needed as static dispatch logic is moved from TensorBody.h, which caused cycles in the first place,
+// to Operators.cpp for supporting multiple backends with multiple kernels.
+//
+// Note [Avoiding Include Cycles In Static Dispatch]
+// In order to avoid #include cycles in the static dispatch build, we've carefully split out
+// the static function definition files into {DispatchKey}Functions.h and {DispatchKey}Functions_inl.h.
+//
+// Without this split, the include cycle looks like TensorBody.h -> CPUFunctions.h -> TensorBody.h.
+// - TensorBody.h #includes CPUFunctions.h in the static dispatch build, because the tensor methods
+//   all need to call into the fastpath C++ API defined in CPUFunctions.h. The methods are also all
+//   directly inlined into TensorBody.h.
+// - CPUFunctions.h #includes TensorBody.h because it contains function declarations for the entire C++ API,
+//   which include functions that have defaultable std::optional<Tensor> arguments.
+//   That requires knowing the full Tensor class definition.
+//
+// We break the cycle by doing the following:
+// - Split out CPUFunction.h into two files: CPUFunctions.h and CPUFunctions_inl.h
+// - CPUFunction.h is a dummy file that just includes the Tensor class and includes CPUFunctions_inl.,
+// - CPUFunctions_inl.h includes everything else
+// - (only in the static dispatch build) TensorBody.h makes sure to finish defining the Tensor class,
+//   and then it includes CPUFunctions_inl.h.
+// - All other files that want the cpu fastpath functions can include CPUFunctions.h directly.
+// - This also means that static dispatch build, CPUFunctions.h only needs to
+//   #include TensorBody.h, and it will automatically bring in CPUFunctions_inl.h.
+#include <ATen/MetaFunctions_inl.h>

.venv/lib/python3.11/site-packages/torch/include/ATen/MethodOperators.h

@@ -0,0 +1,443 @@
+#pragma once
+
+// @generated by torchgen/gen.py from MethodOperators.h
+
+#ifdef TORCH_ASSERT_NO_OPERATORS
+#error This change adds a dependency on native_functions.yaml,             \
+  meaning the file will need to be re-compiled every time an operator      \
+  is changed or added. Consider if your change would be better placed in   \
+  another file, or if a more specific header might achieve the same goal.  \
+  See NOTE: [Tensor vs. TensorBase]
+#endif
+
+// Forward declarations of any types needed in the operator signatures.
+// We can't directly include these classes because it will cause circular include dependencies.
+// This file is included by TensorBody.h, which defines the Tensor class.
+#include <ATen/core/ATen_fwd.h>
+
+#include <ATen/ops/_addmm_activation_ops.h>
+#include <ATen/ops/_autocast_to_full_precision_ops.h>
+#include <ATen/ops/_autocast_to_reduced_precision_ops.h>
+#include <ATen/ops/_backward_ops.h>
+#include <ATen/ops/_coalesced_ops.h>
+#include <ATen/ops/_conj_ops.h>
+#include <ATen/ops/_conj_physical_ops.h>
+#include <ATen/ops/_dimI_ops.h>
+#include <ATen/ops/_dimV_ops.h>
+#include <ATen/ops/_fw_primal_ops.h>
+#include <ATen/ops/_indices_ops.h>
+#include <ATen/ops/_is_all_true_ops.h>
+#include <ATen/ops/_is_any_true_ops.h>
+#include <ATen/ops/_is_zerotensor_ops.h>
+#include <ATen/ops/_lazy_clone_ops.h>
+#include <ATen/ops/_neg_view_ops.h>
+#include <ATen/ops/_nested_tensor_size_ops.h>
+#include <ATen/ops/_nested_tensor_storage_offsets_ops.h>
+#include <ATen/ops/_nested_tensor_strides_ops.h>
+#include <ATen/ops/_nnz_ops.h>
+#include <ATen/ops/_reshape_alias_ops.h>
+#include <ATen/ops/_sparse_mask_projection_ops.h>
+#include <ATen/ops/_to_dense_ops.h>
+#include <ATen/ops/_to_sparse_bsc_ops.h>
+#include <ATen/ops/_to_sparse_bsr_ops.h>
+#include <ATen/ops/_to_sparse_csc_ops.h>
+#include <ATen/ops/_to_sparse_csr_ops.h>
+#include <ATen/ops/_to_sparse_ops.h>
+#include <ATen/ops/_values_ops.h>
+#include <ATen/ops/_version_ops.h>
+#include <ATen/ops/abs_ops.h>
+#include <ATen/ops/absolute_ops.h>
+#include <ATen/ops/acos_ops.h>
+#include <ATen/ops/acosh_ops.h>
+#include <ATen/ops/add_ops.h>
+#include <ATen/ops/addbmm_ops.h>
+#include <ATen/ops/addcdiv_ops.h>
+#include <ATen/ops/addcmul_ops.h>
+#include <ATen/ops/addmm_ops.h>
+#include <ATen/ops/addmv_ops.h>
+#include <ATen/ops/addr_ops.h>
+#include <ATen/ops/adjoint_ops.h>
+#include <ATen/ops/alias_ops.h>
+#include <ATen/ops/align_as_ops.h>
+#include <ATen/ops/align_to_ops.h>
+#include <ATen/ops/all_ops.h>
+#include <ATen/ops/allclose_ops.h>
+#include <ATen/ops/amax_ops.h>
+#include <ATen/ops/amin_ops.h>
+#include <ATen/ops/aminmax_ops.h>
+#include <ATen/ops/and_ops.h>
+#include <ATen/ops/angle_ops.h>
+#include <ATen/ops/any_ops.h>
+#include <ATen/ops/arccos_ops.h>
+#include <ATen/ops/arccosh_ops.h>
+#include <ATen/ops/arcsin_ops.h>
+#include <ATen/ops/arcsinh_ops.h>
+#include <ATen/ops/arctan2_ops.h>
+#include <ATen/ops/arctan_ops.h>
+#include <ATen/ops/arctanh_ops.h>
+#include <ATen/ops/argmax_ops.h>
+#include <ATen/ops/argmin_ops.h>
+#include <ATen/ops/argsort_ops.h>
+#include <ATen/ops/argwhere_ops.h>
+#include <ATen/ops/as_strided_ops.h>
+#include <ATen/ops/as_strided_scatter_ops.h>
+#include <ATen/ops/asin_ops.h>
+#include <ATen/ops/asinh_ops.h>
+#include <ATen/ops/atan2_ops.h>
+#include <ATen/ops/atan_ops.h>
+#include <ATen/ops/atanh_ops.h>
+#include <ATen/ops/baddbmm_ops.h>
+#include <ATen/ops/bernoulli_ops.h>
+#include <ATen/ops/bincount_ops.h>
+#include <ATen/ops/bitwise_and_ops.h>
+#include <ATen/ops/bitwise_left_shift_ops.h>
+#include <ATen/ops/bitwise_not_ops.h>
+#include <ATen/ops/bitwise_or_ops.h>
+#include <ATen/ops/bitwise_right_shift_ops.h>
+#include <ATen/ops/bitwise_xor_ops.h>
+#include <ATen/ops/bmm_ops.h>
+#include <ATen/ops/broadcast_to_ops.h>
+#include <ATen/ops/cauchy_ops.h>
+#include <ATen/ops/ccol_indices_ops.h>
+#include <ATen/ops/ceil_ops.h>
+#include <ATen/ops/chalf_ops.h>
+#include <ATen/ops/cholesky_inverse_ops.h>
+#include <ATen/ops/cholesky_ops.h>
+#include <ATen/ops/cholesky_solve_ops.h>
+#include <ATen/ops/chunk_ops.h>
+#include <ATen/ops/clamp_max_ops.h>
+#include <ATen/ops/clamp_min_ops.h>
+#include <ATen/ops/clamp_ops.h>
+#include <ATen/ops/clip_ops.h>
+#include <ATen/ops/clone_ops.h>
+#include <ATen/ops/coalesce_ops.h>
+#include <ATen/ops/col_indices_ops.h>
+#include <ATen/ops/conj_ops.h>
+#include <ATen/ops/conj_physical_ops.h>
+#include <ATen/ops/contiguous_ops.h>
+#include <ATen/ops/copy_ops.h>
+#include <ATen/ops/copysign_ops.h>
+#include <ATen/ops/corrcoef_ops.h>
+#include <ATen/ops/cos_ops.h>
+#include <ATen/ops/cosh_ops.h>
+#include <ATen/ops/count_nonzero_ops.h>
+#include <ATen/ops/cov_ops.h>
+#include <ATen/ops/cross_ops.h>
+#include <ATen/ops/crow_indices_ops.h>
+#include <ATen/ops/cummax_ops.h>
+#include <ATen/ops/cummin_ops.h>
+#include <ATen/ops/cumprod_ops.h>
+#include <ATen/ops/cumsum_ops.h>
+#include <ATen/ops/data_ops.h>
+#include <ATen/ops/deg2rad_ops.h>
+#include <ATen/ops/dense_dim_ops.h>
+#include <ATen/ops/dequantize_ops.h>
+#include <ATen/ops/det_ops.h>
+#include <ATen/ops/detach_ops.h>
+#include <ATen/ops/diag_embed_ops.h>
+#include <ATen/ops/diag_ops.h>
+#include <ATen/ops/diagflat_ops.h>
+#include <ATen/ops/diagonal_ops.h>
+#include <ATen/ops/diagonal_scatter_ops.h>
+#include <ATen/ops/diff_ops.h>
+#include <ATen/ops/digamma_ops.h>
+#include <ATen/ops/dist_ops.h>
+#include <ATen/ops/div_ops.h>
+#include <ATen/ops/divide_ops.h>
+#include <ATen/ops/dot_ops.h>
+#include <ATen/ops/dsplit_ops.h>
+#include <ATen/ops/eq_ops.h>
+#include <ATen/ops/equal_ops.h>
+#include <ATen/ops/erf_ops.h>
+#include <ATen/ops/erfc_ops.h>
+#include <ATen/ops/erfinv_ops.h>
+#include <ATen/ops/exp2_ops.h>
+#include <ATen/ops/exp_ops.h>
+#include <ATen/ops/expand_as_ops.h>
+#include <ATen/ops/expand_ops.h>
+#include <ATen/ops/expm1_ops.h>
+#include <ATen/ops/exponential_ops.h>
+#include <ATen/ops/fill_diagonal_ops.h>
+#include <ATen/ops/fill_ops.h>
+#include <ATen/ops/fix_ops.h>
+#include <ATen/ops/flatten_ops.h>
+#include <ATen/ops/flip_ops.h>
+#include <ATen/ops/fliplr_ops.h>
+#include <ATen/ops/flipud_ops.h>
+#include <ATen/ops/float_power_ops.h>
+#include <ATen/ops/floor_divide_ops.h>
+#include <ATen/ops/floor_ops.h>
+#include <ATen/ops/fmax_ops.h>
+#include <ATen/ops/fmin_ops.h>
+#include <ATen/ops/fmod_ops.h>
+#include <ATen/ops/frac_ops.h>
+#include <ATen/ops/frexp_ops.h>
+#include <ATen/ops/gather_ops.h>
+#include <ATen/ops/gcd_ops.h>
+#include <ATen/ops/ge_ops.h>
+#include <ATen/ops/geometric_ops.h>
+#include <ATen/ops/geqrf_ops.h>
+#include <ATen/ops/ger_ops.h>
+#include <ATen/ops/greater_equal_ops.h>
+#include <ATen/ops/greater_ops.h>
+#include <ATen/ops/gt_ops.h>
+#include <ATen/ops/hardshrink_backward_ops.h>
+#include <ATen/ops/hardshrink_ops.h>
+#include <ATen/ops/heaviside_ops.h>
+#include <ATen/ops/histc_ops.h>
+#include <ATen/ops/histogram_ops.h>
+#include <ATen/ops/hsplit_ops.h>
+#include <ATen/ops/hypot_ops.h>
+#include <ATen/ops/i0_ops.h>
+#include <ATen/ops/igamma_ops.h>
+#include <ATen/ops/igammac_ops.h>
+#include <ATen/ops/index_add_ops.h>
+#include <ATen/ops/index_copy_ops.h>
+#include <ATen/ops/index_fill_ops.h>
+#include <ATen/ops/index_ops.h>
+#include <ATen/ops/index_put_ops.h>
+#include <ATen/ops/index_reduce_ops.h>
+#include <ATen/ops/index_select_ops.h>
+#include <ATen/ops/indices_ops.h>
+#include <ATen/ops/inner_ops.h>
+#include <ATen/ops/int_repr_ops.h>
+#include <ATen/ops/inverse_ops.h>
+#include <ATen/ops/is_coalesced_ops.h>
+#include <ATen/ops/is_complex_ops.h>
+#include <ATen/ops/is_conj_ops.h>
+#include <ATen/ops/is_distributed_ops.h>
+#include <ATen/ops/is_floating_point_ops.h>
+#include <ATen/ops/is_inference_ops.h>
+#include <ATen/ops/is_leaf_ops.h>
+#include <ATen/ops/is_neg_ops.h>
+#include <ATen/ops/is_nonzero_ops.h>
+#include <ATen/ops/is_pinned_ops.h>
+#include <ATen/ops/is_same_size_ops.h>
+#include <ATen/ops/is_set_to_ops.h>
+#include <ATen/ops/is_signed_ops.h>
+#include <ATen/ops/isclose_ops.h>
+#include <ATen/ops/isfinite_ops.h>
+#include <ATen/ops/isinf_ops.h>
+#include <ATen/ops/isnan_ops.h>
+#include <ATen/ops/isneginf_ops.h>
+#include <ATen/ops/isposinf_ops.h>
+#include <ATen/ops/isreal_ops.h>
+#include <ATen/ops/istft_ops.h>
+#include <ATen/ops/item_ops.h>
+#include <ATen/ops/kron_ops.h>
+#include <ATen/ops/kthvalue_ops.h>
+#include <ATen/ops/lcm_ops.h>
+#include <ATen/ops/ldexp_ops.h>
+#include <ATen/ops/le_ops.h>
+#include <ATen/ops/lerp_ops.h>
+#include <ATen/ops/less_equal_ops.h>
+#include <ATen/ops/less_ops.h>
+#include <ATen/ops/lgamma_ops.h>
+#include <ATen/ops/log10_ops.h>
+#include <ATen/ops/log1p_ops.h>
+#include <ATen/ops/log2_ops.h>
+#include <ATen/ops/log_normal_ops.h>
+#include <ATen/ops/log_ops.h>
+#include <ATen/ops/log_softmax_ops.h>
+#include <ATen/ops/logaddexp2_ops.h>
+#include <ATen/ops/logaddexp_ops.h>
+#include <ATen/ops/logcumsumexp_ops.h>
+#include <ATen/ops/logdet_ops.h>
+#include <ATen/ops/logical_and_ops.h>
+#include <ATen/ops/logical_not_ops.h>
+#include <ATen/ops/logical_or_ops.h>
+#include <ATen/ops/logical_xor_ops.h>
+#include <ATen/ops/logit_ops.h>
+#include <ATen/ops/logsumexp_ops.h>
+#include <ATen/ops/lshift_ops.h>
+#include <ATen/ops/lt_ops.h>
+#include <ATen/ops/lu_solve_ops.h>
+#include <ATen/ops/mH_ops.h>
+#include <ATen/ops/mT_ops.h>
+#include <ATen/ops/masked_fill_ops.h>
+#include <ATen/ops/masked_scatter_ops.h>
+#include <ATen/ops/masked_select_ops.h>
+#include <ATen/ops/matmul_ops.h>
+#include <ATen/ops/matrix_H_ops.h>
+#include <ATen/ops/matrix_exp_ops.h>
+#include <ATen/ops/matrix_power_ops.h>
+#include <ATen/ops/max_ops.h>
+#include <ATen/ops/maximum_ops.h>
+#include <ATen/ops/mean_ops.h>
+#include <ATen/ops/median_ops.h>
+#include <ATen/ops/min_ops.h>
+#include <ATen/ops/minimum_ops.h>
+#include <ATen/ops/mm_ops.h>
+#include <ATen/ops/mode_ops.h>
+#include <ATen/ops/moveaxis_ops.h>
+#include <ATen/ops/movedim_ops.h>
+#include <ATen/ops/msort_ops.h>
+#include <ATen/ops/mul_ops.h>
+#include <ATen/ops/multinomial_ops.h>
+#include <ATen/ops/multiply_ops.h>
+#include <ATen/ops/mv_ops.h>
+#include <ATen/ops/mvlgamma_ops.h>
+#include <ATen/ops/nan_to_num_ops.h>
+#include <ATen/ops/nanmean_ops.h>
+#include <ATen/ops/nanmedian_ops.h>
+#include <ATen/ops/nanquantile_ops.h>
+#include <ATen/ops/nansum_ops.h>
+#include <ATen/ops/narrow_copy_ops.h>
+#include <ATen/ops/narrow_ops.h>
+#include <ATen/ops/ne_ops.h>
+#include <ATen/ops/neg_ops.h>
+#include <ATen/ops/negative_ops.h>
+#include <ATen/ops/new_empty_ops.h>
+#include <ATen/ops/new_empty_strided_ops.h>
+#include <ATen/ops/new_full_ops.h>
+#include <ATen/ops/new_ones_ops.h>
+#include <ATen/ops/new_zeros_ops.h>
+#include <ATen/ops/nextafter_ops.h>
+#include <ATen/ops/nonzero_numpy_ops.h>
+#include <ATen/ops/nonzero_ops.h>
+#include <ATen/ops/nonzero_static_ops.h>
+#include <ATen/ops/norm_ops.h>
+#include <ATen/ops/normal_ops.h>
+#include <ATen/ops/not_equal_ops.h>
+#include <ATen/ops/numpy_T_ops.h>
+#include <ATen/ops/or_ops.h>
+#include <ATen/ops/orgqr_ops.h>
+#include <ATen/ops/ormqr_ops.h>
+#include <ATen/ops/outer_ops.h>
+#include <ATen/ops/output_nr_ops.h>
+#include <ATen/ops/permute_ops.h>
+#include <ATen/ops/pin_memory_ops.h>
+#include <ATen/ops/pinverse_ops.h>
+#include <ATen/ops/polygamma_ops.h>
+#include <ATen/ops/positive_ops.h>
+#include <ATen/ops/pow_ops.h>
+#include <ATen/ops/prelu_ops.h>
+#include <ATen/ops/prod_ops.h>
+#include <ATen/ops/put_ops.h>
+#include <ATen/ops/q_per_channel_axis_ops.h>
+#include <ATen/ops/q_per_channel_scales_ops.h>
+#include <ATen/ops/q_per_channel_zero_points_ops.h>
+#include <ATen/ops/q_scale_ops.h>
+#include <ATen/ops/q_zero_point_ops.h>
+#include <ATen/ops/qr_ops.h>
+#include <ATen/ops/qscheme_ops.h>
+#include <ATen/ops/quantile_ops.h>
+#include <ATen/ops/rad2deg_ops.h>
+#include <ATen/ops/random_ops.h>
+#include <ATen/ops/ravel_ops.h>
+#include <ATen/ops/reciprocal_ops.h>
+#include <ATen/ops/record_stream_ops.h>
+#include <ATen/ops/refine_names_ops.h>
+#include <ATen/ops/relu_ops.h>
+#include <ATen/ops/remainder_ops.h>
+#include <ATen/ops/rename_ops.h>
+#include <ATen/ops/renorm_ops.h>
+#include <ATen/ops/repeat_interleave_ops.h>
+#include <ATen/ops/repeat_ops.h>
+#include <ATen/ops/requires_grad_ops.h>
+#include <ATen/ops/reshape_as_ops.h>
+#include <ATen/ops/reshape_ops.h>
+#include <ATen/ops/resize_as_ops.h>
+#include <ATen/ops/resize_as_sparse_ops.h>
+#include <ATen/ops/resize_ops.h>
+#include <ATen/ops/resolve_conj_ops.h>
+#include <ATen/ops/resolve_neg_ops.h>
+#include <ATen/ops/retain_grad_ops.h>
+#include <ATen/ops/retains_grad_ops.h>
+#include <ATen/ops/roll_ops.h>
+#include <ATen/ops/rot90_ops.h>
+#include <ATen/ops/round_ops.h>
+#include <ATen/ops/row_indices_ops.h>
+#include <ATen/ops/rshift_ops.h>
+#include <ATen/ops/rsqrt_ops.h>
+#include <ATen/ops/scatter_add_ops.h>
+#include <ATen/ops/scatter_ops.h>
+#include <ATen/ops/scatter_reduce_ops.h>
+#include <ATen/ops/select_ops.h>
+#include <ATen/ops/select_scatter_ops.h>
+#include <ATen/ops/set_data_ops.h>
+#include <ATen/ops/set_ops.h>
+#include <ATen/ops/sgn_ops.h>
+#include <ATen/ops/sigmoid_ops.h>
+#include <ATen/ops/sign_ops.h>
+#include <ATen/ops/signbit_ops.h>
+#include <ATen/ops/sin_ops.h>
+#include <ATen/ops/sinc_ops.h>
+#include <ATen/ops/sinh_ops.h>
+#include <ATen/ops/size_ops.h>
+#include <ATen/ops/slice_inverse_ops.h>
+#include <ATen/ops/slice_ops.h>
+#include <ATen/ops/slice_scatter_ops.h>
+#include <ATen/ops/slogdet_ops.h>
+#include <ATen/ops/smm_ops.h>
+#include <ATen/ops/softmax_ops.h>
+#include <ATen/ops/sort_ops.h>
+#include <ATen/ops/sparse_dim_ops.h>
+#include <ATen/ops/sparse_mask_ops.h>
+#include <ATen/ops/sparse_resize_and_clear_ops.h>
+#include <ATen/ops/sparse_resize_ops.h>
+#include <ATen/ops/split_ops.h>
+#include <ATen/ops/split_with_sizes_ops.h>
+#include <ATen/ops/sqrt_ops.h>
+#include <ATen/ops/square_ops.h>
+#include <ATen/ops/squeeze_ops.h>
+#include <ATen/ops/sspaddmm_ops.h>
+#include <ATen/ops/std_ops.h>
+#include <ATen/ops/stft_ops.h>
+#include <ATen/ops/stride_ops.h>
+#include <ATen/ops/sub_ops.h>
+#include <ATen/ops/subtract_ops.h>
+#include <ATen/ops/sum_ops.h>
+#include <ATen/ops/sum_to_size_ops.h>
+#include <ATen/ops/svd_ops.h>
+#include <ATen/ops/swapaxes_ops.h>
+#include <ATen/ops/swapdims_ops.h>
+#include <ATen/ops/t_ops.h>
+#include <ATen/ops/take_along_dim_ops.h>
+#include <ATen/ops/take_ops.h>
+#include <ATen/ops/tan_ops.h>
+#include <ATen/ops/tanh_ops.h>
+#include <ATen/ops/tensor_split_ops.h>
+#include <ATen/ops/tile_ops.h>
+#include <ATen/ops/to_dense_ops.h>
+#include <ATen/ops/to_mkldnn_ops.h>
+#include <ATen/ops/to_ops.h>
+#include <ATen/ops/to_padded_tensor_ops.h>
+#include <ATen/ops/to_sparse_bsc_ops.h>
+#include <ATen/ops/to_sparse_bsr_ops.h>
+#include <ATen/ops/to_sparse_csc_ops.h>
+#include <ATen/ops/to_sparse_csr_ops.h>
+#include <ATen/ops/to_sparse_ops.h>
+#include <ATen/ops/topk_ops.h>
+#include <ATen/ops/trace_ops.h>
+#include <ATen/ops/transpose_ops.h>
+#include <ATen/ops/triangular_solve_ops.h>
+#include <ATen/ops/tril_ops.h>
+#include <ATen/ops/triu_ops.h>
+#include <ATen/ops/true_divide_ops.h>
+#include <ATen/ops/trunc_ops.h>
+#include <ATen/ops/type_as_ops.h>
+#include <ATen/ops/unbind_ops.h>
+#include <ATen/ops/unflatten_ops.h>
+#include <ATen/ops/unfold_ops.h>
+#include <ATen/ops/uniform_ops.h>
+#include <ATen/ops/unsafe_chunk_ops.h>
+#include <ATen/ops/unsafe_split_ops.h>
+#include <ATen/ops/unsafe_split_with_sizes_ops.h>
+#include <ATen/ops/unsqueeze_ops.h>
+#include <ATen/ops/values_ops.h>
+#include <ATen/ops/var_ops.h>
+#include <ATen/ops/vdot_ops.h>
+#include <ATen/ops/view_as_ops.h>
+#include <ATen/ops/view_ops.h>
+#include <ATen/ops/vsplit_ops.h>
+#include <ATen/ops/where_ops.h>
+#include <ATen/ops/xlogy_ops.h>
+#include <ATen/ops/xor_ops.h>
+#include <ATen/ops/zero_ops.h>
+
+namespace at {
+namespace _ops {
+
+} // namespace _ops
+} // namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/NamedTensor.h

@@ -0,0 +1 @@
+#include <ATen/core/NamedTensor.h>

.venv/lib/python3.11/site-packages/torch/include/ATen/NamedTensorUtils.h

@@ -0,0 +1,214 @@
+#pragma once
+#include <ATen/NamedTensor.h>
+#include <ATen/TensorNames.h>
+#include <ATen/WrapDimUtilsMulti.h>
+
+#include <ATen/core/DimVector.h>
+#include <ATen/core/Tensor.h>
+
+namespace at {
+
+using NameVector = SmallVector<Dimname, kDimVectorStaticSize>;
+
+inline bool has_names(const ITensorListRef& tensors) {
+  return std::any_of(tensors.begin(), tensors.end(), [](const Tensor& t) {
+    return t.has_names();
+  });
+}
+
+// Converts dim to an positional index. Errors if `dim` cannot be used to
+// refer to any dimension of tensor.
+TORCH_API int64_t dimname_to_position(const Tensor& tensor, Dimname dim);
+TORCH_API std::vector<int64_t> dimnames_to_positions(
+    const Tensor& tensor,
+    DimnameList dims);
+
+// Unifies two DimnameList to produce a third. This is useful for implementing
+// the named inference rule for binary broadcasting operations like add.
+//
+// There are three main constraints:
+// 1) Check matching: Names must match positionally from the right.
+// 2) Check misaligned: If a name `n` is in `names`, then it must appear at
+//    the same index from the right in other.
+// 3) The output names are obtained by unifying the names individually from the
+// right.
+TORCH_API std::vector<Dimname> unify_from_right(
+    DimnameList names,
+    DimnameList other,
+    const char* action = "broadcast");
+
+[[noreturn]] inline void reportNYIDimnameOverload(const char* op_name) {
+  TORCH_CHECK(
+      false,
+      op_name,
+      ": You passed a dimname (string) to this op in place of a dimension "
+      "index but it does not yet support this behavior. Please pass a dimension "
+      "index to work around this.");
+}
+
+// [NOTE] Writing name inference rules
+//
+// Operators that support named tensors are either composed of operations that
+// support named tensors or implement some name inference rule. An op that
+// implements its own name inference rule generally looks like the following:
+//
+// Tensor op(...) {
+//   perform_shape_checks(...);
+//   # (1)
+//   auto maybe_outnames = compute_outnames(...);
+//   auto result = [&]() {
+//     NoNamesGuard guard;
+//     return op_impl(...);
+//   }();
+//   # (2)
+//   propagate_names_if_nonempty(result, maybe_outnames);
+//
+// Each op has (1) a compute outnames step and (2) a propagate names step.
+//
+// compute_outnames is responsible for checking that input names match and
+// determining what the output names should be. It returns either:
+// - {} (if the inputs tensors are all unnamed)
+// - non-empty outnames.
+//
+// propagate_names_if_nonempty propagates the outnames if they exist to the
+// result tensors.
+//
+// The {} case is an optimization; if the user does not use named tensors they
+// pay no perf cost for it.
+
+namespace namedinference {
+
+const Tensor& propagate_names_if_present_and_nonempty(
+    const Tensor& result,
+    std::optional<DimnameList> maybe_names,
+    bool validate_names = false);
+// Propagates `names` to `result` if `names` is not empty.
+// `names` can be empty; see [NOTE] Writing name inference rules
+// If `names` is not empty, `names.size()` should equal `result.dim()`.
+// When in doubt, use this overload instead of the others.
+TORCH_API const Tensor& propagate_names_if_nonempty(
+    const Tensor& result,
+    DimnameList maybe_names,
+    bool validate_names = false);
+
+// Propagates `names` to `result`. Only use this if we are certain that there
+// are names to propagate (that names is not empty).
+TORCH_API const Tensor& propagate_names(
+    const Tensor& result,
+    DimnameList names,
+    bool validate_names = false);
+
+// Propagates all names from src to result.
+TORCH_API void propagate_names(const Tensor& result, const Tensor& src);
+
+// Propagates all names except for those at the excluded_idxs.
+TORCH_API void propagate_names_except(
+    const Tensor& result,
+    const Tensor& src,
+    IntArrayRef excluded_idxs);
+
+// Used for reduction ops that have a `keepdim` arg.
+TORCH_API void propagate_names_for_reduction(
+    const Tensor& result,
+    const Tensor& src,
+    IntArrayRef excluded_idxs,
+    bool keepdim);
+
+TORCH_API void propagate_names_for_expand(
+    const Tensor& result,
+    const Tensor& self);
+
+TORCH_API std::vector<Dimname> compute_cat_outnames(
+    const MaterializedITensorListRef& tensors);
+
+TORCH_API std::vector<Dimname> compute_broadcast_outnames(
+    const Tensor& self,
+    const Tensor& other);
+
+TORCH_API std::vector<Dimname> broadcast_to_outnames(
+    const Tensor& tensor,
+    const Tensor& reference_tensor,
+    const char* op_name);
+
+TORCH_API std::vector<Dimname> compute_matmul_outnames(
+    const Tensor& self,
+    const Tensor& other);
+
+TORCH_API std::vector<Dimname> compute_cdist_outnames(
+    const Tensor& self,
+    const Tensor& other);
+
+TORCH_API std::vector<Dimname> compute_bmm_outnames(
+    const Tensor& result,
+    const Tensor& self,
+    const Tensor& other);
+
+TORCH_API std::vector<Dimname> compute_squeeze_outnames(const Tensor& tensor);
+TORCH_API std::vector<Dimname> compute_squeeze_outnames(
+    const Tensor& tensor,
+    std::bitset<dim_bitset_size> dims);
+
+std::vector<Dimname> compute_diagonal_outnames(
+    const Tensor& tensor,
+    int64_t dim1,
+    int64_t dim2);
+
+// TensorImpl* overloads for Legacy TH/THC code. Use these sparingly.
+
+TORCH_API TensorImpl* propagate_names_if_nonempty(
+    TensorImpl* result,
+    DimnameList maybe_names,
+    bool validate_names = false);
+
+TORCH_API TensorImpl* propagate_names(
+    TensorImpl* result,
+    DimnameList names,
+    bool validate_names = false);
+
+TORCH_API void propagate_names(TensorImpl* result, /*const */ TensorImpl* src);
+
+TORCH_API inline void propagate_names(
+    const TensorBase& result,
+    DimnameList names,
+    bool validate_names = false) {
+  propagate_names(result.unsafeGetTensorImpl(), names, validate_names);
+}
+
+TORCH_API inline void propagate_names_if_nonempty(
+    const TensorBase& result,
+    DimnameList names,
+    bool validate_names = false) {
+  propagate_names_if_nonempty(
+      result.unsafeGetTensorImpl(), names, validate_names);
+}
+
+TORCH_API inline void propagate_names(
+    const TensorBase& result,
+    const TensorBase& src) {
+  propagate_names(result.unsafeGetTensorImpl(), src.unsafeGetTensorImpl());
+}
+
+// result = m1 @ m2 + bias
+TORCH_API std::vector<Dimname> propagate_names_for_addmm(
+    const Tensor& m1,
+    const Tensor& m2,
+    const Tensor& bias);
+
+TORCH_API std::vector<Dimname> propagate_names_for_addmv(
+    const Tensor& mat,
+    const Tensor& vec,
+    const Tensor& bias);
+
+TORCH_API void check_names_for_dot(TensorImpl* vec1, TensorImpl* vec2);
+
+TORCH_API std::vector<Dimname> compute_baddbmm_outnames(
+    const Tensor& result,
+    const Tensor& self,
+    const Tensor& other,
+    const Tensor& bias);
+
+TORCH_API bool are_names_equal(TensorImpl* self, TensorImpl* other);
+
+} // namespace namedinference
+
+} // namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/NativeFunctions.h

@@ -0,0 +1,1344 @@
+#pragma once
+
+// @generated by torchgen/gen.py from NativeFunctions.h
+
+#ifdef TORCH_ASSERT_NO_OPERATORS
+#error This change adds a dependency on native_functions.yaml,            \
+  meaning the file will need to be re-compiled every time an operator     \
+  is changed or added. Consider if your change would be better placed in  \
+  another file, or if a more specific header might achieve the same goal. \
+  See NOTE: [Tensor vs. TensorBase]
+#endif
+
+#if defined(AT_PER_OPERATOR_HEADERS) && defined(TORCH_ASSERT_ONLY_METHOD_OPERATORS)
+#error This change adds a dependency on all pytorch operators, meaning the      \
+  file will need to be re-compiled every time an operator is changed or added.  \
+  Consider including a specific operator from <ATen/ops/{my_operator}_native.h> \
+  and see NOTE [TORCH_ASSERT_ONLY_METHOD_OPERATORS].
+#endif
+
+#include <c10/core/Scalar.h>
+#include <c10/core/Storage.h>
+#include <c10/core/TensorOptions.h>
+#include <c10/util/Deprecated.h>
+#include <optional>
+#include <c10/core/QScheme.h>
+#include <ATen/core/Reduction.h>
+#include <ATen/core/Tensor.h>
+#include <tuple>
+#include <vector>
+
+#include <ATen/ops/_adaptive_avg_pool2d_native.h>
+#include <ATen/ops/_adaptive_avg_pool2d_backward_native.h>
+#include <ATen/ops/_adaptive_avg_pool3d_native.h>
+#include <ATen/ops/_adaptive_avg_pool3d_backward_native.h>
+#include <ATen/ops/_add_batch_dim_native.h>
+#include <ATen/ops/_add_relu_native.h>
+#include <ATen/ops/_addmm_activation_native.h>
+#include <ATen/ops/_aminmax_native.h>
+#include <ATen/ops/_amp_foreach_non_finite_check_and_unscale_native.h>
+#include <ATen/ops/_amp_update_scale_native.h>
+#include <ATen/ops/_assert_async_native.h>
+#include <ATen/ops/_assert_scalar_native.h>
+#include <ATen/ops/_assert_tensor_metadata_native.h>
+#include <ATen/ops/_autocast_to_full_precision_native.h>
+#include <ATen/ops/_autocast_to_reduced_precision_native.h>
+#include <ATen/ops/_backward_native.h>
+#include <ATen/ops/_batch_norm_impl_index_native.h>
+#include <ATen/ops/_batch_norm_impl_index_backward_native.h>
+#include <ATen/ops/_batch_norm_no_update_native.h>
+#include <ATen/ops/_batch_norm_with_update_native.h>
+#include <ATen/ops/_cast_Byte_native.h>
+#include <ATen/ops/_cast_Char_native.h>
+#include <ATen/ops/_cast_Double_native.h>
+#include <ATen/ops/_cast_Float_native.h>
+#include <ATen/ops/_cast_Half_native.h>
+#include <ATen/ops/_cast_Int_native.h>
+#include <ATen/ops/_cast_Long_native.h>
+#include <ATen/ops/_cast_Short_native.h>
+#include <ATen/ops/_cdist_backward_native.h>
+#include <ATen/ops/_cdist_forward_native.h>
+#include <ATen/ops/_cholesky_solve_helper_native.h>
+#include <ATen/ops/_choose_qparams_per_tensor_native.h>
+#include <ATen/ops/_chunk_cat_native.h>
+#include <ATen/ops/_coalesce_native.h>
+#include <ATen/ops/_coalesced_native.h>
+#include <ATen/ops/_compute_linear_combination_native.h>
+#include <ATen/ops/_conj_native.h>
+#include <ATen/ops/_conj_copy_native.h>
+#include <ATen/ops/_conj_physical_native.h>
+#include <ATen/ops/_conv_depthwise2d_native.h>
+#include <ATen/ops/_convert_indices_from_coo_to_csr_native.h>
+#include <ATen/ops/_convert_indices_from_csr_to_coo_native.h>
+#include <ATen/ops/_convert_weight_to_int4pack_native.h>
+#include <ATen/ops/_convolution_native.h>
+#include <ATen/ops/_convolution_double_backward_native.h>
+#include <ATen/ops/_convolution_mode_native.h>
+#include <ATen/ops/_copy_from_native.h>
+#include <ATen/ops/_copy_from_and_resize_native.h>
+#include <ATen/ops/_cslt_compress_native.h>
+#include <ATen/ops/_cslt_sparse_mm_native.h>
+#include <ATen/ops/_cslt_sparse_mm_search_native.h>
+#include <ATen/ops/_ctc_loss_native.h>
+#include <ATen/ops/_ctc_loss_backward_native.h>
+#include <ATen/ops/_cudnn_ctc_loss_native.h>
+#include <ATen/ops/_cudnn_init_dropout_state_native.h>
+#include <ATen/ops/_cudnn_rnn_native.h>
+#include <ATen/ops/_cudnn_rnn_backward_native.h>
+#include <ATen/ops/_cudnn_rnn_flatten_weight_native.h>
+#include <ATen/ops/_cufft_clear_plan_cache_native.h>
+#include <ATen/ops/_cufft_get_plan_cache_max_size_native.h>
+#include <ATen/ops/_cufft_get_plan_cache_size_native.h>
+#include <ATen/ops/_cufft_set_plan_cache_max_size_native.h>
+#include <ATen/ops/_cummax_helper_native.h>
+#include <ATen/ops/_cummin_helper_native.h>
+#include <ATen/ops/_debug_has_internal_overlap_native.h>
+#include <ATen/ops/_dimI_native.h>
+#include <ATen/ops/_dimV_native.h>
+#include <ATen/ops/_dim_arange_native.h>
+#include <ATen/ops/_dirichlet_grad_native.h>
+#include <ATen/ops/_efficient_attention_backward_native.h>
+#include <ATen/ops/_efficient_attention_forward_native.h>
+#include <ATen/ops/_efficientzerotensor_native.h>
+#include <ATen/ops/_embedding_bag_native.h>
+#include <ATen/ops/_embedding_bag_backward_native.h>
+#include <ATen/ops/_embedding_bag_dense_backward_native.h>
+#include <ATen/ops/_embedding_bag_forward_only_native.h>
+#include <ATen/ops/_embedding_bag_per_sample_weights_backward_native.h>
+#include <ATen/ops/_embedding_bag_sparse_backward_native.h>
+#include <ATen/ops/_empty_affine_quantized_native.h>
+#include <ATen/ops/_empty_per_channel_affine_quantized_native.h>
+#include <ATen/ops/_euclidean_dist_native.h>
+#include <ATen/ops/_fake_quantize_learnable_per_channel_affine_native.h>
+#include <ATen/ops/_fake_quantize_learnable_per_channel_affine_backward_native.h>
+#include <ATen/ops/_fake_quantize_learnable_per_tensor_affine_native.h>
+#include <ATen/ops/_fake_quantize_learnable_per_tensor_affine_backward_native.h>
+#include <ATen/ops/_fake_quantize_per_tensor_affine_cachemask_tensor_qparams_native.h>
+#include <ATen/ops/_fft_c2c_native.h>
+#include <ATen/ops/_fft_c2r_native.h>
+#include <ATen/ops/_fft_r2c_native.h>
+#include <ATen/ops/_fill_mem_eff_dropout_mask_native.h>
+#include <ATen/ops/_flash_attention_backward_native.h>
+#include <ATen/ops/_flash_attention_forward_native.h>
+#include <ATen/ops/_foobar_native.h>
+#include <ATen/ops/_foreach_abs_native.h>
+#include <ATen/ops/_foreach_acos_native.h>
+#include <ATen/ops/_foreach_add_native.h>
+#include <ATen/ops/_foreach_addcdiv_native.h>
+#include <ATen/ops/_foreach_addcmul_native.h>
+#include <ATen/ops/_foreach_asin_native.h>
+#include <ATen/ops/_foreach_atan_native.h>
+#include <ATen/ops/_foreach_ceil_native.h>
+#include <ATen/ops/_foreach_clamp_max_native.h>
+#include <ATen/ops/_foreach_clamp_min_native.h>
+#include <ATen/ops/_foreach_copy_native.h>
+#include <ATen/ops/_foreach_cos_native.h>
+#include <ATen/ops/_foreach_cosh_native.h>
+#include <ATen/ops/_foreach_div_native.h>
+#include <ATen/ops/_foreach_erf_native.h>
+#include <ATen/ops/_foreach_erfc_native.h>
+#include <ATen/ops/_foreach_exp_native.h>
+#include <ATen/ops/_foreach_expm1_native.h>
+#include <ATen/ops/_foreach_floor_native.h>
+#include <ATen/ops/_foreach_frac_native.h>
+#include <ATen/ops/_foreach_lerp_native.h>
+#include <ATen/ops/_foreach_lgamma_native.h>
+#include <ATen/ops/_foreach_log_native.h>
+#include <ATen/ops/_foreach_log10_native.h>
+#include <ATen/ops/_foreach_log1p_native.h>
+#include <ATen/ops/_foreach_log2_native.h>
+#include <ATen/ops/_foreach_max_native.h>
+#include <ATen/ops/_foreach_maximum_native.h>
+#include <ATen/ops/_foreach_minimum_native.h>
+#include <ATen/ops/_foreach_mul_native.h>
+#include <ATen/ops/_foreach_neg_native.h>
+#include <ATen/ops/_foreach_norm_native.h>
+#include <ATen/ops/_foreach_pow_native.h>
+#include <ATen/ops/_foreach_reciprocal_native.h>
+#include <ATen/ops/_foreach_round_native.h>
+#include <ATen/ops/_foreach_sigmoid_native.h>
+#include <ATen/ops/_foreach_sign_native.h>
+#include <ATen/ops/_foreach_sin_native.h>
+#include <ATen/ops/_foreach_sinh_native.h>
+#include <ATen/ops/_foreach_sqrt_native.h>
+#include <ATen/ops/_foreach_sub_native.h>
+#include <ATen/ops/_foreach_tan_native.h>
+#include <ATen/ops/_foreach_tanh_native.h>
+#include <ATen/ops/_foreach_trunc_native.h>
+#include <ATen/ops/_foreach_zero_native.h>
+#include <ATen/ops/_functional_assert_async_native.h>
+#include <ATen/ops/_functional_assert_scalar_native.h>
+#include <ATen/ops/_functional_sym_constrain_range_native.h>
+#include <ATen/ops/_functional_sym_constrain_range_for_size_native.h>
+#include <ATen/ops/_fused_adagrad_native.h>
+#include <ATen/ops/_fused_adam_native.h>
+#include <ATen/ops/_fused_adamw_native.h>
+#include <ATen/ops/_fused_dropout_native.h>
+#include <ATen/ops/_fused_moving_avg_obs_fq_helper_native.h>
+#include <ATen/ops/_fused_sdp_choice_native.h>
+#include <ATen/ops/_fused_sgd_native.h>
+#include <ATen/ops/_fw_primal_native.h>
+#include <ATen/ops/_fw_primal_copy_native.h>
+#include <ATen/ops/_gather_sparse_backward_native.h>
+#include <ATen/ops/_grid_sampler_2d_cpu_fallback_native.h>
+#include <ATen/ops/_grid_sampler_2d_cpu_fallback_backward_native.h>
+#include <ATen/ops/_has_compatible_shallow_copy_type_native.h>
+#include <ATen/ops/_has_same_storage_numel_native.h>
+#include <ATen/ops/_histogramdd_bin_edges_native.h>
+#include <ATen/ops/_histogramdd_from_bin_cts_native.h>
+#include <ATen/ops/_histogramdd_from_bin_tensors_native.h>
+#include <ATen/ops/_index_put_impl_native.h>
+#include <ATen/ops/_indices_native.h>
+#include <ATen/ops/_indices_copy_native.h>
+#include <ATen/ops/_int_mm_native.h>
+#include <ATen/ops/_is_all_true_native.h>
+#include <ATen/ops/_is_any_true_native.h>
+#include <ATen/ops/_is_zerotensor_native.h>
+#include <ATen/ops/_jagged_to_padded_dense_forward_native.h>
+#include <ATen/ops/_lazy_clone_native.h>
+#include <ATen/ops/_linalg_check_errors_native.h>
+#include <ATen/ops/_linalg_det_native.h>
+#include <ATen/ops/_linalg_eigh_native.h>
+#include <ATen/ops/_linalg_eigvals_native.h>
+#include <ATen/ops/_linalg_slogdet_native.h>
+#include <ATen/ops/_linalg_solve_ex_native.h>
+#include <ATen/ops/_linalg_svd_native.h>
+#include <ATen/ops/_local_scalar_dense_native.h>
+#include <ATen/ops/_log_softmax_native.h>
+#include <ATen/ops/_log_softmax_backward_data_native.h>
+#include <ATen/ops/_logcumsumexp_native.h>
+#include <ATen/ops/_lstm_mps_native.h>
+#include <ATen/ops/_lu_with_info_native.h>
+#include <ATen/ops/_make_dep_token_native.h>
+#include <ATen/ops/_make_dual_native.h>
+#include <ATen/ops/_make_dual_copy_native.h>
+#include <ATen/ops/_make_per_channel_quantized_tensor_native.h>
+#include <ATen/ops/_make_per_tensor_quantized_tensor_native.h>
+#include <ATen/ops/_masked_scale_native.h>
+#include <ATen/ops/_masked_softmax_native.h>
+#include <ATen/ops/_masked_softmax_backward_native.h>
+#include <ATen/ops/_mixed_dtypes_linear_native.h>
+#include <ATen/ops/_mkldnn_reshape_native.h>
+#include <ATen/ops/_mkldnn_transpose_native.h>
+#include <ATen/ops/_mps_convolution_native.h>
+#include <ATen/ops/_mps_convolution_transpose_native.h>
+#include <ATen/ops/_native_batch_norm_legit_native.h>
+#include <ATen/ops/_native_batch_norm_legit_no_training_native.h>
+#include <ATen/ops/_native_multi_head_attention_native.h>
+#include <ATen/ops/_neg_view_native.h>
+#include <ATen/ops/_neg_view_copy_native.h>
+#include <ATen/ops/_nested_compute_contiguous_strides_offsets_native.h>
+#include <ATen/ops/_nested_from_padded_native.h>
+#include <ATen/ops/_nested_from_padded_and_nested_example_native.h>
+#include <ATen/ops/_nested_get_jagged_dummy_native.h>
+#include <ATen/ops/_nested_get_lengths_native.h>
+#include <ATen/ops/_nested_get_max_seqlen_native.h>
+#include <ATen/ops/_nested_get_min_seqlen_native.h>
+#include <ATen/ops/_nested_get_offsets_native.h>
+#include <ATen/ops/_nested_get_ragged_idx_native.h>
+#include <ATen/ops/_nested_get_values_native.h>
+#include <ATen/ops/_nested_get_values_copy_native.h>
+#include <ATen/ops/_nested_select_backward_native.h>
+#include <ATen/ops/_nested_sum_backward_native.h>
+#include <ATen/ops/_nested_tensor_from_mask_native.h>
+#include <ATen/ops/_nested_tensor_from_mask_left_aligned_native.h>
+#include <ATen/ops/_nested_tensor_from_tensor_list_native.h>
+#include <ATen/ops/_nested_tensor_size_native.h>
+#include <ATen/ops/_nested_tensor_softmax_with_shape_native.h>
+#include <ATen/ops/_nested_tensor_storage_offsets_native.h>
+#include <ATen/ops/_nested_tensor_strides_native.h>
+#include <ATen/ops/_nested_view_from_buffer_native.h>
+#include <ATen/ops/_nested_view_from_buffer_copy_native.h>
+#include <ATen/ops/_nested_view_from_jagged_native.h>
+#include <ATen/ops/_nested_view_from_jagged_copy_native.h>
+#include <ATen/ops/_new_zeros_with_same_feature_meta_native.h>
+#include <ATen/ops/_nnpack_available_native.h>
+#include <ATen/ops/_nnpack_spatial_convolution_native.h>
+#include <ATen/ops/_nnz_native.h>
+#include <ATen/ops/_pack_padded_sequence_native.h>
+#include <ATen/ops/_pack_padded_sequence_backward_native.h>
+#include <ATen/ops/_pad_circular_native.h>
+#include <ATen/ops/_pad_enum_native.h>
+#include <ATen/ops/_pad_packed_sequence_native.h>
+#include <ATen/ops/_padded_dense_to_jagged_forward_native.h>
+#include <ATen/ops/_pdist_backward_native.h>
+#include <ATen/ops/_pdist_forward_native.h>
+#include <ATen/ops/_pin_memory_native.h>
+#include <ATen/ops/_prelu_kernel_native.h>
+#include <ATen/ops/_prelu_kernel_backward_native.h>
+#include <ATen/ops/_print_native.h>
+#include <ATen/ops/_propagate_xla_data_native.h>
+#include <ATen/ops/_remove_batch_dim_native.h>
+#include <ATen/ops/_reshape_alias_native.h>
+#include <ATen/ops/_reshape_alias_copy_native.h>
+#include <ATen/ops/_reshape_copy_native.h>
+#include <ATen/ops/_reshape_from_tensor_native.h>
+#include <ATen/ops/_resize_output_native.h>
+#include <ATen/ops/_rowwise_prune_native.h>
+#include <ATen/ops/_safe_softmax_native.h>
+#include <ATen/ops/_sample_dirichlet_native.h>
+#include <ATen/ops/_saturate_weight_to_fp16_native.h>
+#include <ATen/ops/_scaled_dot_product_attention_math_native.h>
+#include <ATen/ops/_scaled_dot_product_attention_math_for_mps_native.h>
+#include <ATen/ops/_scaled_dot_product_cudnn_attention_native.h>
+#include <ATen/ops/_scaled_dot_product_cudnn_attention_backward_native.h>
+#include <ATen/ops/_scaled_dot_product_efficient_attention_native.h>
+#include <ATen/ops/_scaled_dot_product_efficient_attention_backward_native.h>
+#include <ATen/ops/_scaled_dot_product_flash_attention_native.h>
+#include <ATen/ops/_scaled_dot_product_flash_attention_backward_native.h>
+#include <ATen/ops/_scaled_dot_product_flash_attention_for_cpu_native.h>
+#include <ATen/ops/_scaled_dot_product_flash_attention_for_cpu_backward_native.h>
+#include <ATen/ops/_scaled_dot_product_fused_attention_overrideable_native.h>
+#include <ATen/ops/_scaled_dot_product_fused_attention_overrideable_backward_native.h>
+#include <ATen/ops/_scaled_mm_native.h>
+#include <ATen/ops/_segment_reduce_backward_native.h>
+#include <ATen/ops/_shape_as_tensor_native.h>
+#include <ATen/ops/_slow_conv2d_backward_native.h>
+#include <ATen/ops/_slow_conv2d_forward_native.h>
+#include <ATen/ops/_sobol_engine_draw_native.h>
+#include <ATen/ops/_sobol_engine_ff_native.h>
+#include <ATen/ops/_sobol_engine_initialize_state_native.h>
+#include <ATen/ops/_sobol_engine_scramble_native.h>
+#include <ATen/ops/_softmax_native.h>
+#include <ATen/ops/_softmax_backward_data_native.h>
+#include <ATen/ops/_sparse_addmm_native.h>
+#include <ATen/ops/_sparse_broadcast_to_native.h>
+#include <ATen/ops/_sparse_broadcast_to_copy_native.h>
+#include <ATen/ops/_sparse_bsc_tensor_unsafe_native.h>
+#include <ATen/ops/_sparse_bsr_tensor_unsafe_native.h>
+#include <ATen/ops/_sparse_compressed_tensor_unsafe_native.h>
+#include <ATen/ops/_sparse_compressed_tensor_with_dims_native.h>
+#include <ATen/ops/_sparse_coo_tensor_unsafe_native.h>
+#include <ATen/ops/_sparse_coo_tensor_with_dims_native.h>
+#include <ATen/ops/_sparse_coo_tensor_with_dims_and_tensors_native.h>
+#include <ATen/ops/_sparse_csc_tensor_unsafe_native.h>
+#include <ATen/ops/_sparse_csr_prod_native.h>
+#include <ATen/ops/_sparse_csr_sum_native.h>
+#include <ATen/ops/_sparse_csr_tensor_unsafe_native.h>
+#include <ATen/ops/_sparse_log_softmax_native.h>
+#include <ATen/ops/_sparse_log_softmax_backward_data_native.h>
+#include <ATen/ops/_sparse_mask_projection_native.h>
+#include <ATen/ops/_sparse_mm_native.h>
+#include <ATen/ops/_sparse_mm_reduce_impl_native.h>
+#include <ATen/ops/_sparse_mm_reduce_impl_backward_native.h>
+#include <ATen/ops/_sparse_semi_structured_addmm_native.h>
+#include <ATen/ops/_sparse_semi_structured_apply_native.h>
+#include <ATen/ops/_sparse_semi_structured_apply_dense_native.h>
+#include <ATen/ops/_sparse_semi_structured_linear_native.h>
+#include <ATen/ops/_sparse_semi_structured_mm_native.h>
+#include <ATen/ops/_sparse_semi_structured_tile_native.h>
+#include <ATen/ops/_sparse_softmax_native.h>
+#include <ATen/ops/_sparse_softmax_backward_data_native.h>
+#include <ATen/ops/_sparse_sparse_matmul_native.h>
+#include <ATen/ops/_sparse_sum_native.h>
+#include <ATen/ops/_sparse_sum_backward_native.h>
+#include <ATen/ops/_spdiags_native.h>
+#include <ATen/ops/_spsolve_native.h>
+#include <ATen/ops/_stack_native.h>
+#include <ATen/ops/_standard_gamma_native.h>
+#include <ATen/ops/_standard_gamma_grad_native.h>
+#include <ATen/ops/_test_ambiguous_defaults_native.h>
+#include <ATen/ops/_test_autograd_multiple_dispatch_native.h>
+#include <ATen/ops/_test_autograd_multiple_dispatch_view_native.h>
+#include <ATen/ops/_test_autograd_multiple_dispatch_view_copy_native.h>
+#include <ATen/ops/_test_check_tensor_native.h>
+#include <ATen/ops/_test_functorch_fallback_native.h>
+#include <ATen/ops/_test_optional_filled_intlist_native.h>
+#include <ATen/ops/_test_optional_floatlist_native.h>
+#include <ATen/ops/_test_optional_intlist_native.h>
+#include <ATen/ops/_test_parallel_materialize_native.h>
+#include <ATen/ops/_test_serialization_subcmul_native.h>
+#include <ATen/ops/_test_string_default_native.h>
+#include <ATen/ops/_test_warn_in_autograd_native.h>
+#include <ATen/ops/_thnn_differentiable_gru_cell_backward_native.h>
+#include <ATen/ops/_thnn_differentiable_lstm_cell_backward_native.h>
+#include <ATen/ops/_thnn_fused_gru_cell_native.h>
+#include <ATen/ops/_thnn_fused_gru_cell_backward_native.h>
+#include <ATen/ops/_thnn_fused_lstm_cell_native.h>
+#include <ATen/ops/_thnn_fused_lstm_cell_backward_native.h>
+#include <ATen/ops/_thnn_fused_lstm_cell_backward_impl_native.h>
+#include <ATen/ops/_to_copy_native.h>
+#include <ATen/ops/_to_cpu_native.h>
+#include <ATen/ops/_to_dense_native.h>
+#include <ATen/ops/_to_sparse_native.h>
+#include <ATen/ops/_to_sparse_bsc_native.h>
+#include <ATen/ops/_to_sparse_bsr_native.h>
+#include <ATen/ops/_to_sparse_csc_native.h>
+#include <ATen/ops/_to_sparse_csr_native.h>
+#include <ATen/ops/_to_sparse_semi_structured_native.h>
+#include <ATen/ops/_transform_bias_rescale_qkv_native.h>
+#include <ATen/ops/_transformer_encoder_layer_fwd_native.h>
+#include <ATen/ops/_trilinear_native.h>
+#include <ATen/ops/_triton_multi_head_attention_native.h>
+#include <ATen/ops/_triton_scaled_dot_attention_native.h>
+#include <ATen/ops/_unique_native.h>
+#include <ATen/ops/_unique2_native.h>
+#include <ATen/ops/_unpack_dual_native.h>
+#include <ATen/ops/_unsafe_index_native.h>
+#include <ATen/ops/_unsafe_index_put_native.h>
+#include <ATen/ops/_unsafe_masked_index_native.h>
+#include <ATen/ops/_unsafe_masked_index_put_accumulate_native.h>
+#include <ATen/ops/_unsafe_view_native.h>
+#include <ATen/ops/_upsample_bicubic2d_aa_native.h>
+#include <ATen/ops/_upsample_bicubic2d_aa_backward_native.h>
+#include <ATen/ops/_upsample_bilinear2d_aa_native.h>
+#include <ATen/ops/_upsample_bilinear2d_aa_backward_native.h>
+#include <ATen/ops/_upsample_nearest_exact1d_native.h>
+#include <ATen/ops/_upsample_nearest_exact1d_backward_native.h>
+#include <ATen/ops/_upsample_nearest_exact2d_native.h>
+#include <ATen/ops/_upsample_nearest_exact2d_backward_native.h>
+#include <ATen/ops/_upsample_nearest_exact3d_native.h>
+#include <ATen/ops/_upsample_nearest_exact3d_backward_native.h>
+#include <ATen/ops/_use_cudnn_ctc_loss_native.h>
+#include <ATen/ops/_use_cudnn_rnn_flatten_weight_native.h>
+#include <ATen/ops/_validate_compressed_sparse_indices_native.h>
+#include <ATen/ops/_validate_sparse_bsc_tensor_args_native.h>
+#include <ATen/ops/_validate_sparse_bsr_tensor_args_native.h>
+#include <ATen/ops/_validate_sparse_compressed_tensor_args_native.h>
+#include <ATen/ops/_validate_sparse_coo_tensor_args_native.h>
+#include <ATen/ops/_validate_sparse_csc_tensor_args_native.h>
+#include <ATen/ops/_validate_sparse_csr_tensor_args_native.h>
+#include <ATen/ops/_values_native.h>
+#include <ATen/ops/_values_copy_native.h>
+#include <ATen/ops/_version_native.h>
+#include <ATen/ops/_weight_int4pack_mm_native.h>
+#include <ATen/ops/_weight_int8pack_mm_native.h>
+#include <ATen/ops/_weight_norm_native.h>
+#include <ATen/ops/_weight_norm_differentiable_backward_native.h>
+#include <ATen/ops/_weight_norm_interface_native.h>
+#include <ATen/ops/_weight_norm_interface_backward_native.h>
+#include <ATen/ops/_wrapped_linear_prepack_native.h>
+#include <ATen/ops/_wrapped_quantized_linear_prepacked_native.h>
+#include <ATen/ops/abs_native.h>
+#include <ATen/ops/absolute_native.h>
+#include <ATen/ops/acos_native.h>
+#include <ATen/ops/acosh_native.h>
+#include <ATen/ops/adaptive_avg_pool1d_native.h>
+#include <ATen/ops/adaptive_avg_pool2d_native.h>
+#include <ATen/ops/adaptive_avg_pool3d_native.h>
+#include <ATen/ops/adaptive_avg_pool3d_backward_native.h>
+#include <ATen/ops/adaptive_max_pool1d_native.h>
+#include <ATen/ops/adaptive_max_pool2d_native.h>
+#include <ATen/ops/adaptive_max_pool2d_backward_native.h>
+#include <ATen/ops/adaptive_max_pool3d_native.h>
+#include <ATen/ops/adaptive_max_pool3d_backward_native.h>
+#include <ATen/ops/add_native.h>
+#include <ATen/ops/addbmm_native.h>
+#include <ATen/ops/addcdiv_native.h>
+#include <ATen/ops/addcmul_native.h>
+#include <ATen/ops/addmm_native.h>
+#include <ATen/ops/addmv_native.h>
+#include <ATen/ops/addr_native.h>
+#include <ATen/ops/adjoint_native.h>
+#include <ATen/ops/affine_grid_generator_native.h>
+#include <ATen/ops/affine_grid_generator_backward_native.h>
+#include <ATen/ops/alias_native.h>
+#include <ATen/ops/alias_copy_native.h>
+#include <ATen/ops/align_as_native.h>
+#include <ATen/ops/align_tensors_native.h>
+#include <ATen/ops/align_to_native.h>
+#include <ATen/ops/all_native.h>
+#include <ATen/ops/allclose_native.h>
+#include <ATen/ops/alpha_dropout_native.h>
+#include <ATen/ops/amax_native.h>
+#include <ATen/ops/amin_native.h>
+#include <ATen/ops/aminmax_native.h>
+#include <ATen/ops/and_native.h>
+#include <ATen/ops/angle_native.h>
+#include <ATen/ops/any_native.h>
+#include <ATen/ops/arange_native.h>
+#include <ATen/ops/arccos_native.h>
+#include <ATen/ops/arccosh_native.h>
+#include <ATen/ops/arcsin_native.h>
+#include <ATen/ops/arcsinh_native.h>
+#include <ATen/ops/arctan_native.h>
+#include <ATen/ops/arctan2_native.h>
+#include <ATen/ops/arctanh_native.h>
+#include <ATen/ops/argmax_native.h>
+#include <ATen/ops/argmin_native.h>
+#include <ATen/ops/argsort_native.h>
+#include <ATen/ops/argwhere_native.h>
+#include <ATen/ops/as_strided_native.h>
+#include <ATen/ops/as_strided_copy_native.h>
+#include <ATen/ops/as_strided_scatter_native.h>
+#include <ATen/ops/asin_native.h>
+#include <ATen/ops/asinh_native.h>
+#include <ATen/ops/atan_native.h>
+#include <ATen/ops/atan2_native.h>
+#include <ATen/ops/atanh_native.h>
+#include <ATen/ops/atleast_1d_native.h>
+#include <ATen/ops/atleast_2d_native.h>
+#include <ATen/ops/atleast_3d_native.h>
+#include <ATen/ops/avg_pool1d_native.h>
+#include <ATen/ops/avg_pool2d_native.h>
+#include <ATen/ops/avg_pool2d_backward_native.h>
+#include <ATen/ops/avg_pool3d_native.h>
+#include <ATen/ops/avg_pool3d_backward_native.h>
+#include <ATen/ops/baddbmm_native.h>
+#include <ATen/ops/bartlett_window_native.h>
+#include <ATen/ops/batch_norm_native.h>
+#include <ATen/ops/batch_norm_backward_native.h>
+#include <ATen/ops/batch_norm_backward_elemt_native.h>
+#include <ATen/ops/batch_norm_backward_reduce_native.h>
+#include <ATen/ops/batch_norm_elemt_native.h>
+#include <ATen/ops/batch_norm_gather_stats_native.h>
+#include <ATen/ops/batch_norm_gather_stats_with_counts_native.h>
+#include <ATen/ops/batch_norm_stats_native.h>
+#include <ATen/ops/batch_norm_update_stats_native.h>
+#include <ATen/ops/bernoulli_native.h>
+#include <ATen/ops/bilinear_native.h>
+#include <ATen/ops/binary_cross_entropy_native.h>
+#include <ATen/ops/binary_cross_entropy_backward_native.h>
+#include <ATen/ops/binary_cross_entropy_with_logits_native.h>
+#include <ATen/ops/bincount_native.h>
+#include <ATen/ops/binomial_native.h>
+#include <ATen/ops/bitwise_and_native.h>
+#include <ATen/ops/bitwise_left_shift_native.h>
+#include <ATen/ops/bitwise_not_native.h>
+#include <ATen/ops/bitwise_or_native.h>
+#include <ATen/ops/bitwise_right_shift_native.h>
+#include <ATen/ops/bitwise_xor_native.h>
+#include <ATen/ops/blackman_window_native.h>
+#include <ATen/ops/block_diag_native.h>
+#include <ATen/ops/bmm_native.h>
+#include <ATen/ops/broadcast_tensors_native.h>
+#include <ATen/ops/broadcast_to_native.h>
+#include <ATen/ops/bucketize_native.h>
+#include <ATen/ops/can_cast_native.h>
+#include <ATen/ops/cartesian_prod_native.h>
+#include <ATen/ops/cat_native.h>
+#include <ATen/ops/cauchy_native.h>
+#include <ATen/ops/ccol_indices_native.h>
+#include <ATen/ops/ccol_indices_copy_native.h>
+#include <ATen/ops/cdist_native.h>
+#include <ATen/ops/ceil_native.h>
+#include <ATen/ops/celu_native.h>
+#include <ATen/ops/chain_matmul_native.h>
+#include <ATen/ops/chalf_native.h>
+#include <ATen/ops/channel_shuffle_native.h>
+#include <ATen/ops/cholesky_native.h>
+#include <ATen/ops/cholesky_inverse_native.h>
+#include <ATen/ops/cholesky_solve_native.h>
+#include <ATen/ops/choose_qparams_optimized_native.h>
+#include <ATen/ops/chunk_native.h>
+#include <ATen/ops/clamp_native.h>
+#include <ATen/ops/clamp_max_native.h>
+#include <ATen/ops/clamp_min_native.h>
+#include <ATen/ops/clip_native.h>
+#include <ATen/ops/clone_native.h>
+#include <ATen/ops/coalesce_native.h>
+#include <ATen/ops/col2im_native.h>
+#include <ATen/ops/col_indices_native.h>
+#include <ATen/ops/col_indices_copy_native.h>
+#include <ATen/ops/column_stack_native.h>
+#include <ATen/ops/combinations_native.h>
+#include <ATen/ops/complex_native.h>
+#include <ATen/ops/concat_native.h>
+#include <ATen/ops/concatenate_native.h>
+#include <ATen/ops/conj_native.h>
+#include <ATen/ops/conj_physical_native.h>
+#include <ATen/ops/constant_pad_nd_native.h>
+#include <ATen/ops/contiguous_native.h>
+#include <ATen/ops/conv1d_native.h>
+#include <ATen/ops/conv2d_native.h>
+#include <ATen/ops/conv3d_native.h>
+#include <ATen/ops/conv_depthwise3d_native.h>
+#include <ATen/ops/conv_tbc_native.h>
+#include <ATen/ops/conv_tbc_backward_native.h>
+#include <ATen/ops/conv_transpose1d_native.h>
+#include <ATen/ops/conv_transpose2d_native.h>
+#include <ATen/ops/conv_transpose3d_native.h>
+#include <ATen/ops/convolution_native.h>
+#include <ATen/ops/convolution_backward_native.h>
+#include <ATen/ops/convolution_backward_overrideable_native.h>
+#include <ATen/ops/convolution_overrideable_native.h>
+#include <ATen/ops/copy_native.h>
+#include <ATen/ops/copy_sparse_to_sparse_native.h>
+#include <ATen/ops/copysign_native.h>
+#include <ATen/ops/corrcoef_native.h>
+#include <ATen/ops/cos_native.h>
+#include <ATen/ops/cosh_native.h>
+#include <ATen/ops/cosine_embedding_loss_native.h>
+#include <ATen/ops/cosine_similarity_native.h>
+#include <ATen/ops/count_nonzero_native.h>
+#include <ATen/ops/cov_native.h>
+#include <ATen/ops/cross_native.h>
+#include <ATen/ops/cross_entropy_loss_native.h>
+#include <ATen/ops/crow_indices_native.h>
+#include <ATen/ops/crow_indices_copy_native.h>
+#include <ATen/ops/ctc_loss_native.h>
+#include <ATen/ops/cudnn_affine_grid_generator_native.h>
+#include <ATen/ops/cudnn_affine_grid_generator_backward_native.h>
+#include <ATen/ops/cudnn_batch_norm_native.h>
+#include <ATen/ops/cudnn_batch_norm_backward_native.h>
+#include <ATen/ops/cudnn_convolution_native.h>
+#include <ATen/ops/cudnn_convolution_add_relu_native.h>
+#include <ATen/ops/cudnn_convolution_relu_native.h>
+#include <ATen/ops/cudnn_convolution_transpose_native.h>
+#include <ATen/ops/cudnn_grid_sampler_native.h>
+#include <ATen/ops/cudnn_grid_sampler_backward_native.h>
+#include <ATen/ops/cudnn_is_acceptable_native.h>
+#include <ATen/ops/cummax_native.h>
+#include <ATen/ops/cummaxmin_backward_native.h>
+#include <ATen/ops/cummin_native.h>
+#include <ATen/ops/cumprod_native.h>
+#include <ATen/ops/cumprod_backward_native.h>
+#include <ATen/ops/cumsum_native.h>
+#include <ATen/ops/cumulative_trapezoid_native.h>
+#include <ATen/ops/data_native.h>
+#include <ATen/ops/deg2rad_native.h>
+#include <ATen/ops/dense_dim_native.h>
+#include <ATen/ops/dequantize_native.h>
+#include <ATen/ops/det_native.h>
+#include <ATen/ops/detach_native.h>
+#include <ATen/ops/detach_copy_native.h>
+#include <ATen/ops/diag_native.h>
+#include <ATen/ops/diag_embed_native.h>
+#include <ATen/ops/diagflat_native.h>
+#include <ATen/ops/diagonal_native.h>
+#include <ATen/ops/diagonal_backward_native.h>
+#include <ATen/ops/diagonal_copy_native.h>
+#include <ATen/ops/diagonal_scatter_native.h>
+#include <ATen/ops/diff_native.h>
+#include <ATen/ops/digamma_native.h>
+#include <ATen/ops/dist_native.h>
+#include <ATen/ops/div_native.h>
+#include <ATen/ops/divide_native.h>
+#include <ATen/ops/dot_native.h>
+#include <ATen/ops/dropout_native.h>
+#include <ATen/ops/dsplit_native.h>
+#include <ATen/ops/dstack_native.h>
+#include <ATen/ops/einsum_native.h>
+#include <ATen/ops/elu_native.h>
+#include <ATen/ops/elu_backward_native.h>
+#include <ATen/ops/embedding_native.h>
+#include <ATen/ops/embedding_backward_native.h>
+#include <ATen/ops/embedding_bag_native.h>
+#include <ATen/ops/embedding_dense_backward_native.h>
+#include <ATen/ops/embedding_renorm_native.h>
+#include <ATen/ops/embedding_sparse_backward_native.h>
+#include <ATen/ops/empty_native.h>
+#include <ATen/ops/empty_like_native.h>
+#include <ATen/ops/empty_permuted_native.h>
+#include <ATen/ops/empty_quantized_native.h>
+#include <ATen/ops/empty_strided_native.h>
+#include <ATen/ops/eq_native.h>
+#include <ATen/ops/equal_native.h>
+#include <ATen/ops/erf_native.h>
+#include <ATen/ops/erfc_native.h>
+#include <ATen/ops/erfinv_native.h>
+#include <ATen/ops/exp_native.h>
+#include <ATen/ops/exp2_native.h>
+#include <ATen/ops/expand_native.h>
+#include <ATen/ops/expand_as_native.h>
+#include <ATen/ops/expand_copy_native.h>
+#include <ATen/ops/expm1_native.h>
+#include <ATen/ops/exponential_native.h>
+#include <ATen/ops/eye_native.h>
+#include <ATen/ops/fake_quantize_per_channel_affine_native.h>
+#include <ATen/ops/fake_quantize_per_channel_affine_cachemask_native.h>
+#include <ATen/ops/fake_quantize_per_channel_affine_cachemask_backward_native.h>
+#include <ATen/ops/fake_quantize_per_tensor_affine_native.h>
+#include <ATen/ops/fake_quantize_per_tensor_affine_cachemask_native.h>
+#include <ATen/ops/fake_quantize_per_tensor_affine_cachemask_backward_native.h>
+#include <ATen/ops/fbgemm_linear_fp16_weight_native.h>
+#include <ATen/ops/fbgemm_linear_fp16_weight_fp32_activation_native.h>
+#include <ATen/ops/fbgemm_linear_int8_weight_native.h>
+#include <ATen/ops/fbgemm_linear_int8_weight_fp32_activation_native.h>
+#include <ATen/ops/fbgemm_linear_quantize_weight_native.h>
+#include <ATen/ops/fbgemm_pack_gemm_matrix_fp16_native.h>
+#include <ATen/ops/fbgemm_pack_quantized_matrix_native.h>
+#include <ATen/ops/feature_alpha_dropout_native.h>
+#include <ATen/ops/feature_dropout_native.h>
+#include <ATen/ops/fft_fft_native.h>
+#include <ATen/ops/fft_fft2_native.h>
+#include <ATen/ops/fft_fftfreq_native.h>
+#include <ATen/ops/fft_fftn_native.h>
+#include <ATen/ops/fft_fftshift_native.h>
+#include <ATen/ops/fft_hfft_native.h>
+#include <ATen/ops/fft_hfft2_native.h>
+#include <ATen/ops/fft_hfftn_native.h>
+#include <ATen/ops/fft_ifft_native.h>
+#include <ATen/ops/fft_ifft2_native.h>
+#include <ATen/ops/fft_ifftn_native.h>
+#include <ATen/ops/fft_ifftshift_native.h>
+#include <ATen/ops/fft_ihfft_native.h>
+#include <ATen/ops/fft_ihfft2_native.h>
+#include <ATen/ops/fft_ihfftn_native.h>
+#include <ATen/ops/fft_irfft_native.h>
+#include <ATen/ops/fft_irfft2_native.h>
+#include <ATen/ops/fft_irfftn_native.h>
+#include <ATen/ops/fft_rfft_native.h>
+#include <ATen/ops/fft_rfft2_native.h>
+#include <ATen/ops/fft_rfftfreq_native.h>
+#include <ATen/ops/fft_rfftn_native.h>
+#include <ATen/ops/fill_native.h>
+#include <ATen/ops/fill_diagonal_native.h>
+#include <ATen/ops/fix_native.h>
+#include <ATen/ops/flatten_native.h>
+#include <ATen/ops/flatten_dense_tensors_native.h>
+#include <ATen/ops/flip_native.h>
+#include <ATen/ops/fliplr_native.h>
+#include <ATen/ops/flipud_native.h>
+#include <ATen/ops/float_power_native.h>
+#include <ATen/ops/floor_native.h>
+#include <ATen/ops/floor_divide_native.h>
+#include <ATen/ops/fmax_native.h>
+#include <ATen/ops/fmin_native.h>
+#include <ATen/ops/fmod_native.h>
+#include <ATen/ops/frac_native.h>
+#include <ATen/ops/fractional_max_pool2d_native.h>
+#include <ATen/ops/fractional_max_pool2d_backward_native.h>
+#include <ATen/ops/fractional_max_pool3d_native.h>
+#include <ATen/ops/fractional_max_pool3d_backward_native.h>
+#include <ATen/ops/frexp_native.h>
+#include <ATen/ops/frobenius_norm_native.h>
+#include <ATen/ops/from_file_native.h>
+#include <ATen/ops/full_native.h>
+#include <ATen/ops/full_like_native.h>
+#include <ATen/ops/fused_moving_avg_obs_fake_quant_native.h>
+#include <ATen/ops/gather_native.h>
+#include <ATen/ops/gather_backward_native.h>
+#include <ATen/ops/gcd_native.h>
+#include <ATen/ops/ge_native.h>
+#include <ATen/ops/gelu_native.h>
+#include <ATen/ops/gelu_backward_native.h>
+#include <ATen/ops/geometric_native.h>
+#include <ATen/ops/geqrf_native.h>
+#include <ATen/ops/ger_native.h>
+#include <ATen/ops/glu_native.h>
+#include <ATen/ops/glu_backward_native.h>
+#include <ATen/ops/glu_backward_jvp_native.h>
+#include <ATen/ops/glu_jvp_native.h>
+#include <ATen/ops/gradient_native.h>
+#include <ATen/ops/greater_native.h>
+#include <ATen/ops/greater_equal_native.h>
+#include <ATen/ops/grid_sampler_native.h>
+#include <ATen/ops/grid_sampler_2d_native.h>
+#include <ATen/ops/grid_sampler_2d_backward_native.h>
+#include <ATen/ops/grid_sampler_3d_native.h>
+#include <ATen/ops/grid_sampler_3d_backward_native.h>
+#include <ATen/ops/group_norm_native.h>
+#include <ATen/ops/gru_native.h>
+#include <ATen/ops/gru_cell_native.h>
+#include <ATen/ops/gt_native.h>
+#include <ATen/ops/hamming_window_native.h>
+#include <ATen/ops/hann_window_native.h>
+#include <ATen/ops/hardshrink_native.h>
+#include <ATen/ops/hardshrink_backward_native.h>
+#include <ATen/ops/hardsigmoid_native.h>
+#include <ATen/ops/hardsigmoid_backward_native.h>
+#include <ATen/ops/hardswish_native.h>
+#include <ATen/ops/hardswish_backward_native.h>
+#include <ATen/ops/hardtanh_native.h>
+#include <ATen/ops/hardtanh_backward_native.h>
+#include <ATen/ops/heaviside_native.h>
+#include <ATen/ops/hinge_embedding_loss_native.h>
+#include <ATen/ops/histc_native.h>
+#include <ATen/ops/histogram_native.h>
+#include <ATen/ops/histogramdd_native.h>
+#include <ATen/ops/hsplit_native.h>
+#include <ATen/ops/hspmm_native.h>
+#include <ATen/ops/hstack_native.h>
+#include <ATen/ops/huber_loss_native.h>
+#include <ATen/ops/huber_loss_backward_native.h>
+#include <ATen/ops/hypot_native.h>
+#include <ATen/ops/i0_native.h>
+#include <ATen/ops/igamma_native.h>
+#include <ATen/ops/igammac_native.h>
+#include <ATen/ops/im2col_native.h>
+#include <ATen/ops/imag_native.h>
+#include <ATen/ops/index_native.h>
+#include <ATen/ops/index_add_native.h>
+#include <ATen/ops/index_copy_native.h>
+#include <ATen/ops/index_fill_native.h>
+#include <ATen/ops/index_put_native.h>
+#include <ATen/ops/index_reduce_native.h>
+#include <ATen/ops/index_select_native.h>
+#include <ATen/ops/index_select_backward_native.h>
+#include <ATen/ops/indices_native.h>
+#include <ATen/ops/indices_copy_native.h>
+#include <ATen/ops/infinitely_differentiable_gelu_backward_native.h>
+#include <ATen/ops/inner_native.h>
+#include <ATen/ops/instance_norm_native.h>
+#include <ATen/ops/int_repr_native.h>
+#include <ATen/ops/inverse_native.h>
+#include <ATen/ops/is_coalesced_native.h>
+#include <ATen/ops/is_complex_native.h>
+#include <ATen/ops/is_conj_native.h>
+#include <ATen/ops/is_distributed_native.h>
+#include <ATen/ops/is_floating_point_native.h>
+#include <ATen/ops/is_inference_native.h>
+#include <ATen/ops/is_leaf_native.h>
+#include <ATen/ops/is_neg_native.h>
+#include <ATen/ops/is_nonzero_native.h>
+#include <ATen/ops/is_pinned_native.h>
+#include <ATen/ops/is_same_size_native.h>
+#include <ATen/ops/is_set_to_native.h>
+#include <ATen/ops/is_signed_native.h>
+#include <ATen/ops/is_vulkan_available_native.h>
+#include <ATen/ops/isclose_native.h>
+#include <ATen/ops/isfinite_native.h>
+#include <ATen/ops/isin_native.h>
+#include <ATen/ops/isinf_native.h>
+#include <ATen/ops/isnan_native.h>
+#include <ATen/ops/isneginf_native.h>
+#include <ATen/ops/isposinf_native.h>
+#include <ATen/ops/isreal_native.h>
+#include <ATen/ops/istft_native.h>
+#include <ATen/ops/item_native.h>
+#include <ATen/ops/kaiser_window_native.h>
+#include <ATen/ops/kl_div_native.h>
+#include <ATen/ops/kron_native.h>
+#include <ATen/ops/kthvalue_native.h>
+#include <ATen/ops/l1_loss_native.h>
+#include <ATen/ops/layer_norm_native.h>
+#include <ATen/ops/lcm_native.h>
+#include <ATen/ops/ldexp_native.h>
+#include <ATen/ops/le_native.h>
+#include <ATen/ops/leaky_relu_native.h>
+#include <ATen/ops/leaky_relu_backward_native.h>
+#include <ATen/ops/lerp_native.h>
+#include <ATen/ops/less_native.h>
+#include <ATen/ops/less_equal_native.h>
+#include <ATen/ops/lgamma_native.h>
+#include <ATen/ops/lift_native.h>
+#include <ATen/ops/lift_fresh_native.h>
+#include <ATen/ops/lift_fresh_copy_native.h>
+#include <ATen/ops/linalg_cholesky_native.h>
+#include <ATen/ops/linalg_cholesky_ex_native.h>
+#include <ATen/ops/linalg_cond_native.h>
+#include <ATen/ops/linalg_cross_native.h>
+#include <ATen/ops/linalg_det_native.h>
+#include <ATen/ops/linalg_diagonal_native.h>
+#include <ATen/ops/linalg_eig_native.h>
+#include <ATen/ops/linalg_eigh_native.h>
+#include <ATen/ops/linalg_eigvals_native.h>
+#include <ATen/ops/linalg_eigvalsh_native.h>
+#include <ATen/ops/linalg_householder_product_native.h>
+#include <ATen/ops/linalg_inv_native.h>
+#include <ATen/ops/linalg_inv_ex_native.h>
+#include <ATen/ops/linalg_ldl_factor_native.h>
+#include <ATen/ops/linalg_ldl_factor_ex_native.h>
+#include <ATen/ops/linalg_ldl_solve_native.h>
+#include <ATen/ops/linalg_lstsq_native.h>
+#include <ATen/ops/linalg_lu_native.h>
+#include <ATen/ops/linalg_lu_factor_native.h>
+#include <ATen/ops/linalg_lu_factor_ex_native.h>
+#include <ATen/ops/linalg_lu_solve_native.h>
+#include <ATen/ops/linalg_matmul_native.h>
+#include <ATen/ops/linalg_matrix_exp_native.h>
+#include <ATen/ops/linalg_matrix_norm_native.h>
+#include <ATen/ops/linalg_matrix_power_native.h>
+#include <ATen/ops/linalg_matrix_rank_native.h>
+#include <ATen/ops/linalg_multi_dot_native.h>
+#include <ATen/ops/linalg_norm_native.h>
+#include <ATen/ops/linalg_pinv_native.h>
+#include <ATen/ops/linalg_qr_native.h>
+#include <ATen/ops/linalg_slogdet_native.h>
+#include <ATen/ops/linalg_solve_native.h>
+#include <ATen/ops/linalg_solve_ex_native.h>
+#include <ATen/ops/linalg_solve_triangular_native.h>
+#include <ATen/ops/linalg_svd_native.h>
+#include <ATen/ops/linalg_svdvals_native.h>
+#include <ATen/ops/linalg_tensorinv_native.h>
+#include <ATen/ops/linalg_tensorsolve_native.h>
+#include <ATen/ops/linalg_vander_native.h>
+#include <ATen/ops/linalg_vecdot_native.h>
+#include <ATen/ops/linalg_vector_norm_native.h>
+#include <ATen/ops/linear_native.h>
+#include <ATen/ops/linear_backward_native.h>
+#include <ATen/ops/linspace_native.h>
+#include <ATen/ops/log_native.h>
+#include <ATen/ops/log10_native.h>
+#include <ATen/ops/log1p_native.h>
+#include <ATen/ops/log2_native.h>
+#include <ATen/ops/log_normal_native.h>
+#include <ATen/ops/log_sigmoid_native.h>
+#include <ATen/ops/log_sigmoid_backward_native.h>
+#include <ATen/ops/log_sigmoid_forward_native.h>
+#include <ATen/ops/log_softmax_native.h>
+#include <ATen/ops/logaddexp_native.h>
+#include <ATen/ops/logaddexp2_native.h>
+#include <ATen/ops/logcumsumexp_native.h>
+#include <ATen/ops/logdet_native.h>
+#include <ATen/ops/logical_and_native.h>
+#include <ATen/ops/logical_not_native.h>
+#include <ATen/ops/logical_or_native.h>
+#include <ATen/ops/logical_xor_native.h>
+#include <ATen/ops/logit_native.h>
+#include <ATen/ops/logit_backward_native.h>
+#include <ATen/ops/logspace_native.h>
+#include <ATen/ops/logsumexp_native.h>
+#include <ATen/ops/lshift_native.h>
+#include <ATen/ops/lstm_native.h>
+#include <ATen/ops/lstm_cell_native.h>
+#include <ATen/ops/lstm_mps_backward_native.h>
+#include <ATen/ops/lt_native.h>
+#include <ATen/ops/lu_solve_native.h>
+#include <ATen/ops/lu_unpack_native.h>
+#include <ATen/ops/mH_native.h>
+#include <ATen/ops/mT_native.h>
+#include <ATen/ops/margin_ranking_loss_native.h>
+#include <ATen/ops/masked_fill_native.h>
+#include <ATen/ops/masked_scatter_native.h>
+#include <ATen/ops/masked_scatter_backward_native.h>
+#include <ATen/ops/masked_select_native.h>
+#include <ATen/ops/masked_select_backward_native.h>
+#include <ATen/ops/matmul_native.h>
+#include <ATen/ops/matmul_backward_native.h>
+#include <ATen/ops/matrix_H_native.h>
+#include <ATen/ops/matrix_exp_native.h>
+#include <ATen/ops/matrix_exp_backward_native.h>
+#include <ATen/ops/matrix_power_native.h>
+#include <ATen/ops/max_native.h>
+#include <ATen/ops/max_pool1d_native.h>
+#include <ATen/ops/max_pool1d_with_indices_native.h>
+#include <ATen/ops/max_pool2d_native.h>
+#include <ATen/ops/max_pool2d_backward_native.h>
+#include <ATen/ops/max_pool2d_with_indices_native.h>
+#include <ATen/ops/max_pool2d_with_indices_backward_native.h>
+#include <ATen/ops/max_pool3d_native.h>
+#include <ATen/ops/max_pool3d_with_indices_native.h>
+#include <ATen/ops/max_pool3d_with_indices_backward_native.h>
+#include <ATen/ops/max_unpool2d_native.h>
+#include <ATen/ops/max_unpool3d_native.h>
+#include <ATen/ops/maximum_native.h>
+#include <ATen/ops/mean_native.h>
+#include <ATen/ops/median_native.h>
+#include <ATen/ops/meshgrid_native.h>
+#include <ATen/ops/min_native.h>
+#include <ATen/ops/minimum_native.h>
+#include <ATen/ops/miopen_batch_norm_native.h>
+#include <ATen/ops/miopen_batch_norm_backward_native.h>
+#include <ATen/ops/miopen_convolution_native.h>
+#include <ATen/ops/miopen_convolution_add_relu_native.h>
+#include <ATen/ops/miopen_convolution_relu_native.h>
+#include <ATen/ops/miopen_convolution_transpose_native.h>
+#include <ATen/ops/miopen_depthwise_convolution_native.h>
+#include <ATen/ops/miopen_rnn_native.h>
+#include <ATen/ops/miopen_rnn_backward_native.h>
+#include <ATen/ops/mish_native.h>
+#include <ATen/ops/mish_backward_native.h>
+#include <ATen/ops/mkldnn_adaptive_avg_pool2d_native.h>
+#include <ATen/ops/mkldnn_adaptive_avg_pool2d_backward_native.h>
+#include <ATen/ops/mkldnn_convolution_native.h>
+#include <ATen/ops/mkldnn_linear_native.h>
+#include <ATen/ops/mkldnn_linear_backward_native.h>
+#include <ATen/ops/mkldnn_linear_backward_input_native.h>
+#include <ATen/ops/mkldnn_linear_backward_weights_native.h>
+#include <ATen/ops/mkldnn_max_pool2d_native.h>
+#include <ATen/ops/mkldnn_max_pool2d_backward_native.h>
+#include <ATen/ops/mkldnn_max_pool3d_native.h>
+#include <ATen/ops/mkldnn_max_pool3d_backward_native.h>
+#include <ATen/ops/mkldnn_reorder_conv2d_weight_native.h>
+#include <ATen/ops/mkldnn_reorder_conv3d_weight_native.h>
+#include <ATen/ops/mkldnn_rnn_layer_native.h>
+#include <ATen/ops/mkldnn_rnn_layer_backward_native.h>
+#include <ATen/ops/mm_native.h>
+#include <ATen/ops/mode_native.h>
+#include <ATen/ops/moveaxis_native.h>
+#include <ATen/ops/movedim_native.h>
+#include <ATen/ops/mps_convolution_backward_native.h>
+#include <ATen/ops/mps_convolution_transpose_backward_native.h>
+#include <ATen/ops/mse_loss_native.h>
+#include <ATen/ops/mse_loss_backward_native.h>
+#include <ATen/ops/msort_native.h>
+#include <ATen/ops/mul_native.h>
+#include <ATen/ops/multi_margin_loss_native.h>
+#include <ATen/ops/multi_margin_loss_backward_native.h>
+#include <ATen/ops/multilabel_margin_loss_native.h>
+#include <ATen/ops/multilabel_margin_loss_backward_native.h>
+#include <ATen/ops/multilabel_margin_loss_forward_native.h>
+#include <ATen/ops/multinomial_native.h>
+#include <ATen/ops/multiply_native.h>
+#include <ATen/ops/mv_native.h>
+#include <ATen/ops/mvlgamma_native.h>
+#include <ATen/ops/nan_to_num_native.h>
+#include <ATen/ops/nanmean_native.h>
+#include <ATen/ops/nanmedian_native.h>
+#include <ATen/ops/nanquantile_native.h>
+#include <ATen/ops/nansum_native.h>
+#include <ATen/ops/narrow_native.h>
+#include <ATen/ops/narrow_copy_native.h>
+#include <ATen/ops/native_batch_norm_native.h>
+#include <ATen/ops/native_batch_norm_backward_native.h>
+#include <ATen/ops/native_channel_shuffle_native.h>
+#include <ATen/ops/native_dropout_native.h>
+#include <ATen/ops/native_dropout_backward_native.h>
+#include <ATen/ops/native_group_norm_native.h>
+#include <ATen/ops/native_group_norm_backward_native.h>
+#include <ATen/ops/native_layer_norm_native.h>
+#include <ATen/ops/native_layer_norm_backward_native.h>
+#include <ATen/ops/native_norm_native.h>
+#include <ATen/ops/ne_native.h>
+#include <ATen/ops/neg_native.h>
+#include <ATen/ops/negative_native.h>
+#include <ATen/ops/nested_to_padded_tensor_native.h>
+#include <ATen/ops/new_empty_native.h>
+#include <ATen/ops/new_empty_strided_native.h>
+#include <ATen/ops/new_full_native.h>
+#include <ATen/ops/new_ones_native.h>
+#include <ATen/ops/new_zeros_native.h>
+#include <ATen/ops/nextafter_native.h>
+#include <ATen/ops/nll_loss_native.h>
+#include <ATen/ops/nll_loss2d_native.h>
+#include <ATen/ops/nll_loss2d_backward_native.h>
+#include <ATen/ops/nll_loss2d_forward_native.h>
+#include <ATen/ops/nll_loss_backward_native.h>
+#include <ATen/ops/nll_loss_forward_native.h>
+#include <ATen/ops/nll_loss_nd_native.h>
+#include <ATen/ops/nonzero_native.h>
+#include <ATen/ops/nonzero_numpy_native.h>
+#include <ATen/ops/nonzero_static_native.h>
+#include <ATen/ops/norm_native.h>
+#include <ATen/ops/norm_except_dim_native.h>
+#include <ATen/ops/normal_native.h>
+#include <ATen/ops/not_equal_native.h>
+#include <ATen/ops/nuclear_norm_native.h>
+#include <ATen/ops/numpy_T_native.h>
+#include <ATen/ops/one_hot_native.h>
+#include <ATen/ops/ones_native.h>
+#include <ATen/ops/ones_like_native.h>
+#include <ATen/ops/or_native.h>
+#include <ATen/ops/orgqr_native.h>
+#include <ATen/ops/ormqr_native.h>
+#include <ATen/ops/outer_native.h>
+#include <ATen/ops/output_nr_native.h>
+#include <ATen/ops/pad_native.h>
+#include <ATen/ops/pad_sequence_native.h>
+#include <ATen/ops/pairwise_distance_native.h>
+#include <ATen/ops/pdist_native.h>
+#include <ATen/ops/permute_native.h>
+#include <ATen/ops/permute_copy_native.h>
+#include <ATen/ops/pin_memory_native.h>
+#include <ATen/ops/pinverse_native.h>
+#include <ATen/ops/pixel_shuffle_native.h>
+#include <ATen/ops/pixel_unshuffle_native.h>
+#include <ATen/ops/poisson_native.h>
+#include <ATen/ops/poisson_nll_loss_native.h>
+#include <ATen/ops/polar_native.h>
+#include <ATen/ops/polygamma_native.h>
+#include <ATen/ops/positive_native.h>
+#include <ATen/ops/pow_native.h>
+#include <ATen/ops/prelu_native.h>
+#include <ATen/ops/prod_native.h>
+#include <ATen/ops/promote_types_native.h>
+#include <ATen/ops/put_native.h>
+#include <ATen/ops/q_per_channel_axis_native.h>
+#include <ATen/ops/q_per_channel_scales_native.h>
+#include <ATen/ops/q_per_channel_zero_points_native.h>
+#include <ATen/ops/q_scale_native.h>
+#include <ATen/ops/q_zero_point_native.h>
+#include <ATen/ops/qr_native.h>
+#include <ATen/ops/qscheme_native.h>
+#include <ATen/ops/quantile_native.h>
+#include <ATen/ops/quantize_per_channel_native.h>
+#include <ATen/ops/quantize_per_tensor_native.h>
+#include <ATen/ops/quantize_per_tensor_dynamic_native.h>
+#include <ATen/ops/quantized_batch_norm_native.h>
+#include <ATen/ops/quantized_gru_cell_native.h>
+#include <ATen/ops/quantized_lstm_cell_native.h>
+#include <ATen/ops/quantized_max_pool1d_native.h>
+#include <ATen/ops/quantized_max_pool2d_native.h>
+#include <ATen/ops/quantized_max_pool3d_native.h>
+#include <ATen/ops/quantized_rnn_relu_cell_native.h>
+#include <ATen/ops/quantized_rnn_tanh_cell_native.h>
+#include <ATen/ops/rad2deg_native.h>
+#include <ATen/ops/rand_native.h>
+#include <ATen/ops/rand_like_native.h>
+#include <ATen/ops/randint_native.h>
+#include <ATen/ops/randint_like_native.h>
+#include <ATen/ops/randn_native.h>
+#include <ATen/ops/randn_like_native.h>
+#include <ATen/ops/random_native.h>
+#include <ATen/ops/randperm_native.h>
+#include <ATen/ops/range_native.h>
+#include <ATen/ops/ravel_native.h>
+#include <ATen/ops/real_native.h>
+#include <ATen/ops/reciprocal_native.h>
+#include <ATen/ops/record_stream_native.h>
+#include <ATen/ops/refine_names_native.h>
+#include <ATen/ops/reflection_pad1d_native.h>
+#include <ATen/ops/reflection_pad1d_backward_native.h>
+#include <ATen/ops/reflection_pad2d_native.h>
+#include <ATen/ops/reflection_pad2d_backward_native.h>
+#include <ATen/ops/reflection_pad3d_native.h>
+#include <ATen/ops/reflection_pad3d_backward_native.h>
+#include <ATen/ops/relu_native.h>
+#include <ATen/ops/relu6_native.h>
+#include <ATen/ops/remainder_native.h>
+#include <ATen/ops/rename_native.h>
+#include <ATen/ops/renorm_native.h>
+#include <ATen/ops/repeat_native.h>
+#include <ATen/ops/repeat_interleave_native.h>
+#include <ATen/ops/replication_pad1d_native.h>
+#include <ATen/ops/replication_pad1d_backward_native.h>
+#include <ATen/ops/replication_pad2d_native.h>
+#include <ATen/ops/replication_pad2d_backward_native.h>
+#include <ATen/ops/replication_pad3d_native.h>
+#include <ATen/ops/replication_pad3d_backward_native.h>
+#include <ATen/ops/requires_grad_native.h>
+#include <ATen/ops/reshape_native.h>
+#include <ATen/ops/reshape_as_native.h>
+#include <ATen/ops/resize_native.h>
+#include <ATen/ops/resize_as_native.h>
+#include <ATen/ops/resize_as_sparse_native.h>
+#include <ATen/ops/resolve_conj_native.h>
+#include <ATen/ops/resolve_neg_native.h>
+#include <ATen/ops/result_type_native.h>
+#include <ATen/ops/retain_grad_native.h>
+#include <ATen/ops/retains_grad_native.h>
+#include <ATen/ops/rms_norm_native.h>
+#include <ATen/ops/rnn_relu_native.h>
+#include <ATen/ops/rnn_relu_cell_native.h>
+#include <ATen/ops/rnn_tanh_native.h>
+#include <ATen/ops/rnn_tanh_cell_native.h>
+#include <ATen/ops/roll_native.h>
+#include <ATen/ops/rot90_native.h>
+#include <ATen/ops/round_native.h>
+#include <ATen/ops/row_indices_native.h>
+#include <ATen/ops/row_indices_copy_native.h>
+#include <ATen/ops/row_stack_native.h>
+#include <ATen/ops/rrelu_native.h>
+#include <ATen/ops/rrelu_with_noise_native.h>
+#include <ATen/ops/rrelu_with_noise_backward_native.h>
+#include <ATen/ops/rshift_native.h>
+#include <ATen/ops/rsqrt_native.h>
+#include <ATen/ops/rsub_native.h>
+#include <ATen/ops/scalar_tensor_native.h>
+#include <ATen/ops/scaled_dot_product_attention_native.h>
+#include <ATen/ops/scatter_native.h>
+#include <ATen/ops/scatter_add_native.h>
+#include <ATen/ops/scatter_reduce_native.h>
+#include <ATen/ops/searchsorted_native.h>
+#include <ATen/ops/segment_reduce_native.h>
+#include <ATen/ops/select_native.h>
+#include <ATen/ops/select_backward_native.h>
+#include <ATen/ops/select_copy_native.h>
+#include <ATen/ops/select_scatter_native.h>
+#include <ATen/ops/selu_native.h>
+#include <ATen/ops/set_native.h>
+#include <ATen/ops/set_data_native.h>
+#include <ATen/ops/sgn_native.h>
+#include <ATen/ops/sigmoid_native.h>
+#include <ATen/ops/sigmoid_backward_native.h>
+#include <ATen/ops/sign_native.h>
+#include <ATen/ops/signbit_native.h>
+#include <ATen/ops/silu_native.h>
+#include <ATen/ops/silu_backward_native.h>
+#include <ATen/ops/sin_native.h>
+#include <ATen/ops/sinc_native.h>
+#include <ATen/ops/sinh_native.h>
+#include <ATen/ops/size_native.h>
+#include <ATen/ops/slice_native.h>
+#include <ATen/ops/slice_backward_native.h>
+#include <ATen/ops/slice_copy_native.h>
+#include <ATen/ops/slice_inverse_native.h>
+#include <ATen/ops/slice_scatter_native.h>
+#include <ATen/ops/slogdet_native.h>
+#include <ATen/ops/slow_conv3d_native.h>
+#include <ATen/ops/slow_conv3d_forward_native.h>
+#include <ATen/ops/slow_conv_dilated2d_native.h>
+#include <ATen/ops/slow_conv_dilated3d_native.h>
+#include <ATen/ops/slow_conv_transpose2d_native.h>
+#include <ATen/ops/slow_conv_transpose3d_native.h>
+#include <ATen/ops/smm_native.h>
+#include <ATen/ops/smooth_l1_loss_native.h>
+#include <ATen/ops/smooth_l1_loss_backward_native.h>
+#include <ATen/ops/soft_margin_loss_native.h>
+#include <ATen/ops/soft_margin_loss_backward_native.h>
+#include <ATen/ops/softmax_native.h>
+#include <ATen/ops/softplus_native.h>
+#include <ATen/ops/softplus_backward_native.h>
+#include <ATen/ops/softshrink_native.h>
+#include <ATen/ops/softshrink_backward_native.h>
+#include <ATen/ops/sort_native.h>
+#include <ATen/ops/sparse_bsc_tensor_native.h>
+#include <ATen/ops/sparse_bsr_tensor_native.h>
+#include <ATen/ops/sparse_compressed_tensor_native.h>
+#include <ATen/ops/sparse_coo_tensor_native.h>
+#include <ATen/ops/sparse_csc_tensor_native.h>
+#include <ATen/ops/sparse_csr_tensor_native.h>
+#include <ATen/ops/sparse_dim_native.h>
+#include <ATen/ops/sparse_mask_native.h>
+#include <ATen/ops/sparse_resize_native.h>
+#include <ATen/ops/sparse_resize_and_clear_native.h>
+#include <ATen/ops/sparse_sampled_addmm_native.h>
+#include <ATen/ops/special_airy_ai_native.h>
+#include <ATen/ops/special_bessel_j0_native.h>
+#include <ATen/ops/special_bessel_j1_native.h>
+#include <ATen/ops/special_bessel_y0_native.h>
+#include <ATen/ops/special_bessel_y1_native.h>
+#include <ATen/ops/special_chebyshev_polynomial_t_native.h>
+#include <ATen/ops/special_chebyshev_polynomial_u_native.h>
+#include <ATen/ops/special_chebyshev_polynomial_v_native.h>
+#include <ATen/ops/special_chebyshev_polynomial_w_native.h>
+#include <ATen/ops/special_digamma_native.h>
+#include <ATen/ops/special_entr_native.h>
+#include <ATen/ops/special_erf_native.h>
+#include <ATen/ops/special_erfc_native.h>
+#include <ATen/ops/special_erfcx_native.h>
+#include <ATen/ops/special_erfinv_native.h>
+#include <ATen/ops/special_exp2_native.h>
+#include <ATen/ops/special_expit_native.h>
+#include <ATen/ops/special_expm1_native.h>
+#include <ATen/ops/special_gammainc_native.h>
+#include <ATen/ops/special_gammaincc_native.h>
+#include <ATen/ops/special_gammaln_native.h>
+#include <ATen/ops/special_hermite_polynomial_h_native.h>
+#include <ATen/ops/special_hermite_polynomial_he_native.h>
+#include <ATen/ops/special_i0_native.h>
+#include <ATen/ops/special_i0e_native.h>
+#include <ATen/ops/special_i1_native.h>
+#include <ATen/ops/special_i1e_native.h>
+#include <ATen/ops/special_laguerre_polynomial_l_native.h>
+#include <ATen/ops/special_legendre_polynomial_p_native.h>
+#include <ATen/ops/special_log1p_native.h>
+#include <ATen/ops/special_log_ndtr_native.h>
+#include <ATen/ops/special_log_softmax_native.h>
+#include <ATen/ops/special_logit_native.h>
+#include <ATen/ops/special_logsumexp_native.h>
+#include <ATen/ops/special_modified_bessel_i0_native.h>
+#include <ATen/ops/special_modified_bessel_i1_native.h>
+#include <ATen/ops/special_modified_bessel_k0_native.h>
+#include <ATen/ops/special_modified_bessel_k1_native.h>
+#include <ATen/ops/special_multigammaln_native.h>
+#include <ATen/ops/special_ndtr_native.h>
+#include <ATen/ops/special_ndtri_native.h>
+#include <ATen/ops/special_polygamma_native.h>
+#include <ATen/ops/special_psi_native.h>
+#include <ATen/ops/special_round_native.h>
+#include <ATen/ops/special_scaled_modified_bessel_k0_native.h>
+#include <ATen/ops/special_scaled_modified_bessel_k1_native.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_t_native.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_u_native.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_v_native.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_w_native.h>
+#include <ATen/ops/special_sinc_native.h>
+#include <ATen/ops/special_softmax_native.h>
+#include <ATen/ops/special_spherical_bessel_j0_native.h>
+#include <ATen/ops/special_xlog1py_native.h>
+#include <ATen/ops/special_xlogy_native.h>
+#include <ATen/ops/special_zeta_native.h>
+#include <ATen/ops/split_native.h>
+#include <ATen/ops/split_copy_native.h>
+#include <ATen/ops/split_with_sizes_native.h>
+#include <ATen/ops/split_with_sizes_copy_native.h>
+#include <ATen/ops/sqrt_native.h>
+#include <ATen/ops/square_native.h>
+#include <ATen/ops/squeeze_native.h>
+#include <ATen/ops/squeeze_copy_native.h>
+#include <ATen/ops/sspaddmm_native.h>
+#include <ATen/ops/stack_native.h>
+#include <ATen/ops/std_native.h>
+#include <ATen/ops/std_mean_native.h>
+#include <ATen/ops/stft_native.h>
+#include <ATen/ops/stride_native.h>
+#include <ATen/ops/sub_native.h>
+#include <ATen/ops/subtract_native.h>
+#include <ATen/ops/sum_native.h>
+#include <ATen/ops/sum_to_size_native.h>
+#include <ATen/ops/svd_native.h>
+#include <ATen/ops/swapaxes_native.h>
+#include <ATen/ops/swapdims_native.h>
+#include <ATen/ops/sym_constrain_range_native.h>
+#include <ATen/ops/sym_constrain_range_for_size_native.h>
+#include <ATen/ops/sym_numel_native.h>
+#include <ATen/ops/sym_size_native.h>
+#include <ATen/ops/sym_storage_offset_native.h>
+#include <ATen/ops/sym_stride_native.h>
+#include <ATen/ops/t_native.h>
+#include <ATen/ops/t_copy_native.h>
+#include <ATen/ops/take_native.h>
+#include <ATen/ops/take_along_dim_native.h>
+#include <ATen/ops/tan_native.h>
+#include <ATen/ops/tanh_native.h>
+#include <ATen/ops/tanh_backward_native.h>
+#include <ATen/ops/tensor_split_native.h>
+#include <ATen/ops/tensordot_native.h>
+#include <ATen/ops/thnn_conv2d_native.h>
+#include <ATen/ops/threshold_native.h>
+#include <ATen/ops/threshold_backward_native.h>
+#include <ATen/ops/tile_native.h>
+#include <ATen/ops/to_native.h>
+#include <ATen/ops/to_dense_native.h>
+#include <ATen/ops/to_dense_backward_native.h>
+#include <ATen/ops/to_mkldnn_native.h>
+#include <ATen/ops/to_mkldnn_backward_native.h>
+#include <ATen/ops/to_padded_tensor_native.h>
+#include <ATen/ops/to_sparse_native.h>
+#include <ATen/ops/to_sparse_bsc_native.h>
+#include <ATen/ops/to_sparse_bsr_native.h>
+#include <ATen/ops/to_sparse_csc_native.h>
+#include <ATen/ops/to_sparse_csr_native.h>
+#include <ATen/ops/topk_native.h>
+#include <ATen/ops/trace_native.h>
+#include <ATen/ops/trace_backward_native.h>
+#include <ATen/ops/transpose_native.h>
+#include <ATen/ops/transpose_copy_native.h>
+#include <ATen/ops/trapezoid_native.h>
+#include <ATen/ops/trapz_native.h>
+#include <ATen/ops/triangular_solve_native.h>
+#include <ATen/ops/tril_native.h>
+#include <ATen/ops/tril_indices_native.h>
+#include <ATen/ops/triplet_margin_loss_native.h>
+#include <ATen/ops/triu_native.h>
+#include <ATen/ops/triu_indices_native.h>
+#include <ATen/ops/true_divide_native.h>
+#include <ATen/ops/trunc_native.h>
+#include <ATen/ops/type_as_native.h>
+#include <ATen/ops/unbind_native.h>
+#include <ATen/ops/unbind_copy_native.h>
+#include <ATen/ops/unflatten_native.h>
+#include <ATen/ops/unflatten_dense_tensors_native.h>
+#include <ATen/ops/unfold_native.h>
+#include <ATen/ops/unfold_backward_native.h>
+#include <ATen/ops/unfold_copy_native.h>
+#include <ATen/ops/uniform_native.h>
+#include <ATen/ops/unique_consecutive_native.h>
+#include <ATen/ops/unique_dim_native.h>
+#include <ATen/ops/unique_dim_consecutive_native.h>
+#include <ATen/ops/unsafe_chunk_native.h>
+#include <ATen/ops/unsafe_split_native.h>
+#include <ATen/ops/unsafe_split_with_sizes_native.h>
+#include <ATen/ops/unsqueeze_native.h>
+#include <ATen/ops/unsqueeze_copy_native.h>
+#include <ATen/ops/upsample_bicubic2d_native.h>
+#include <ATen/ops/upsample_bicubic2d_backward_native.h>
+#include <ATen/ops/upsample_bilinear2d_native.h>
+#include <ATen/ops/upsample_bilinear2d_backward_native.h>
+#include <ATen/ops/upsample_linear1d_native.h>
+#include <ATen/ops/upsample_linear1d_backward_native.h>
+#include <ATen/ops/upsample_nearest1d_native.h>
+#include <ATen/ops/upsample_nearest1d_backward_native.h>
+#include <ATen/ops/upsample_nearest2d_native.h>
+#include <ATen/ops/upsample_nearest2d_backward_native.h>
+#include <ATen/ops/upsample_nearest3d_native.h>
+#include <ATen/ops/upsample_nearest3d_backward_native.h>
+#include <ATen/ops/upsample_trilinear3d_native.h>
+#include <ATen/ops/upsample_trilinear3d_backward_native.h>
+#include <ATen/ops/value_selecting_reduction_backward_native.h>
+#include <ATen/ops/values_native.h>
+#include <ATen/ops/values_copy_native.h>
+#include <ATen/ops/vander_native.h>
+#include <ATen/ops/var_native.h>
+#include <ATen/ops/var_mean_native.h>
+#include <ATen/ops/vdot_native.h>
+#include <ATen/ops/view_native.h>
+#include <ATen/ops/view_as_native.h>
+#include <ATen/ops/view_as_complex_native.h>
+#include <ATen/ops/view_as_complex_copy_native.h>
+#include <ATen/ops/view_as_real_native.h>
+#include <ATen/ops/view_as_real_copy_native.h>
+#include <ATen/ops/view_copy_native.h>
+#include <ATen/ops/vsplit_native.h>
+#include <ATen/ops/vstack_native.h>
+#include <ATen/ops/where_native.h>
+#include <ATen/ops/xlogy_native.h>
+#include <ATen/ops/xor_native.h>
+#include <ATen/ops/zero_native.h>
+#include <ATen/ops/zeros_native.h>
+#include <ATen/ops/zeros_like_native.h>
+
+

.venv/lib/python3.11/site-packages/torch/include/ATen/NestedTensorImpl.h

@@ -0,0 +1,286 @@
+#pragma once
+#include <ATen/MemoryOverlap.h>
+#include <ATen/Tensor.h>
+#include <c10/core/DispatchKey.h>
+#include <c10/core/DispatchKeySet.h>
+#include <c10/core/MemoryFormat.h>
+#include <c10/core/TensorImpl.h>
+#include <c10/util/ArrayRef.h>
+#include <c10/util/Exception.h>
+#include <c10/util/Metaprogramming.h>
+#include <c10/util/irange.h>
+
+namespace at::native {
+struct NestedTensorImpl;
+inline bool nested_tensor_impl_is_contiguous(const NestedTensorImpl* nt);
+int64_t get_numel_from_nested_size_tensor(const at::Tensor& tensor);
+at::Tensor construct_nested_strides(const at::Tensor& nested_size);
+at::Tensor construct_offsets(const at::Tensor& nested_size);
+
+struct TORCH_API NestedTensorImpl : public c10::TensorImpl {
+  explicit NestedTensorImpl(
+      Storage storage,
+      c10::DispatchKeySet key_set,
+      const caffe2::TypeMeta data_type,
+      at::Tensor nested_sizes,
+      at::Tensor nested_strides,
+      at::Tensor storage_offsets);
+
+  explicit NestedTensorImpl(
+      const at::Tensor& buffer,
+      at::Tensor nested_sizes,
+      at::Tensor nested_strides,
+      at::Tensor storage_offsets);
+  // assume contiguous, `nested_strides` and `offsets`
+  // can be infered from `nested_sizes`
+  explicit NestedTensorImpl(
+      const at::Tensor& buffer,
+      const at::Tensor& nested_sizes);
+
+  // This constructor is used creating view tensors from nested tensors
+  explicit NestedTensorImpl(
+      c10::TensorImpl::ImplType impl_type,
+      const at::Tensor& base_tensor,
+      at::Tensor nested_sizes,
+      at::Tensor nested_strides,
+      at::Tensor storage_offsets);
+
+  // TODO: don't expose private implementation details like this; in
+  // particular, resizing this tensor will mess up our dim() and
+  // callers cannot fix it.
+  const Tensor& get_nested_sizes() const {
+    return nested_sizes_;
+  }
+  // TODO: don't expose private implementation details like this
+  const Tensor& get_nested_strides() const {
+    return nested_strides_;
+  }
+  const Tensor& get_storage_offsets() const {
+    return storage_offsets_;
+  }
+  // Returns nullopt if the ith dimension is irregular. The ith dimension
+  // of a NestedTensor is regular if the unbound tensors match in
+  // size at the (i-1)th dimension.
+  std::optional<int64_t> opt_size(int64_t d) const;
+
+  int64_t size(int64_t d) const {
+    std::optional<int64_t> optional_size = this->opt_size(d);
+    TORCH_CHECK(
+        optional_size.has_value(),
+        "Given dimension ",
+        d,
+        " is irregular and does not have a size.");
+    return *optional_size;
+  }
+  /**
+   * Return a view of the nested tensor as a 1 dimensional contiguous tensor.
+   *
+   * The buffer tensor created by this function shares the same storage_impl as
+   * the original nested tensor, and therefore can be seen as a view.
+   *
+   * @return A newly constructed view tensor
+   */
+  at::Tensor get_buffer() const {
+    TORCH_CHECK(
+        nested_tensor_impl_is_contiguous(this),
+        "NestedTensor must be contiguous to get buffer.");
+    return get_unsafe_storage_as_tensor();
+  }
+  /**
+   * If possible use get_buffer() instead. This function returns the storage
+   * as a tensor directly, which is not safe to use in general. If using this
+   * function, The caller must ensure to account for nested_sizes,
+   * nested_strides and storage_offsets.
+   *
+   * @return A newly constructed view tensor
+   */
+  at::Tensor get_unsafe_storage_as_tensor() const {
+    auto buffer_key_set_ = generate_buffer_key_set();
+    const auto buffer_size = get_buffer_size();
+    auto buffer_tensor_impl = c10::make_intrusive<TensorImpl>(
+        c10::TensorImpl::VIEW, Storage(storage_), buffer_key_set_, data_type_);
+    buffer_tensor_impl->set_sizes_contiguous(
+        c10::makeArrayRef(static_cast<int64_t>(buffer_size)));
+    return Tensor(buffer_tensor_impl);
+  }
+
+  size_t get_buffer_size() const {
+    return storage_.nbytes() / data_type_.itemsize();
+  }
+
+ protected:
+  const char* tensorimpl_type_name() const override;
+
+  // TODO: numel_custom and is_contiguous_custom can be profitably overridden
+  // with real implementations
+  int64_t numel_custom() const override;
+  c10::SymInt sym_numel_custom() const override;
+  bool is_contiguous_custom(MemoryFormat) const override;
+  int64_t size_custom(int64_t d) const override {
+    return this->size(d);
+  }
+  c10::SymInt sym_size_custom(int64_t d) const override {
+    return c10::SymInt{this->size(d)};
+  }
+  IntArrayRef sizes_custom() const override;
+  c10::SymIntArrayRef sym_sizes_custom() const override;
+  IntArrayRef strides_custom() const override;
+  c10::SymIntArrayRef sym_strides_custom() const override;
+
+  // this one is real
+  int64_t dim_custom() const override;
+
+  c10::intrusive_ptr<TensorImpl> shallow_copy_and_detach(
+      const c10::VariableVersion& version_counter,
+      bool allow_tensor_metadata_change) const override;
+
+  c10::intrusive_ptr<TensorImpl> shallow_copy_and_detach(
+      c10::VariableVersion&& version_counter,
+      bool allow_tensor_metadata_change) const override;
+
+  void shallow_copy_from(const c10::intrusive_ptr<TensorImpl>& impl) override {
+    copy_tensor_metadata(
+        /*src_impl=*/impl.get(),
+        /*dest_impl=*/this,
+        /*version_counter=*/version_counter(),
+        /*allow_tensor_metadata_change=*/allow_tensor_metadata_change());
+  }
+
+ private:
+  // Must be called after any changes to our dim() to sync the state
+  // to TensorImpl.
+  void refresh_dim();
+
+  // NOLINTNEXTLINE(cppcoreguidelines-avoid-const-or-ref-data-members)
+  const at::Tensor nested_sizes_, nested_strides_;
+  // The starting positions of the underlying tensors in contiguous buffer
+  // i.e. the buffer memory offsets to get the underlying tensors
+  // The reason to keep this metadata is that, without strong enough constraint
+  // it cannot be derived from `nested_sizes_`
+  // and `nested_strides_`:
+  // 1. when buffer has blanks, e.g. [tensor1, blank, tensor2]
+  //    this can happen e.g. after slicing a nested tensor
+  // 2. when multiple tensors share a same memory
+  // 3. when the nesting ordering is changed, e.g. [tensor1, tensor3, tensor2]
+  // Some strong enough constraints are:
+  // 1. every underlying tensor is contiguous in memory
+  //    && nesting in ascending order
+  // NOLINTNEXTLINE(cppcoreguidelines-avoid-const-or-ref-data-members)
+  const at::Tensor storage_offsets_;
+  // NOTE: -1 here means the size is missing
+  // Optional to allow it to be computed lazily from nested.
+  // TODO: maybe we can remove this metadata since
+  //       we can compute it from `nested_sizes_`
+  mutable std::optional<std::vector<int64_t>> opt_sizes_;
+
+  template <typename VariableVersion>
+  c10::intrusive_ptr<TensorImpl> shallow_copy_and_detach_core(
+      VariableVersion&& version_counter,
+      bool allow_tensor_metadata_change) const;
+
+  /**
+   * Generates a non-nested key_set from a nested tensor.
+   *
+   * For many nested tensor kernel implementations a buffer tensor
+   * is generated and redispatched to a non-nested kernel this function
+   * generates the key set used by that buffer tensor
+   *
+   * @return Appropriate key set for non-nested tensor
+   */
+  inline c10::DispatchKeySet generate_buffer_key_set() const {
+    auto buffer_key_set = this->key_set();
+    const bool Autograd = buffer_key_set.has_any(c10::autograd_dispatch_keyset);
+    // Remove nested tensor specific keys
+    buffer_key_set = buffer_key_set -
+        c10::DispatchKeySet{
+            c10::DispatchKey::NestedTensor,
+            c10::DispatchKey::AutogradNestedTensor};
+
+    // Add dense tensor specific keys
+    buffer_key_set =
+        buffer_key_set | c10::DispatchKeySet{c10::DispatchKey::Dense};
+    buffer_key_set = Autograd
+        ? c10::DispatchKeySet{c10::DispatchKey::Autograd} | buffer_key_set
+        : buffer_key_set;
+
+    return buffer_key_set;
+  }
+};
+
+inline NestedTensorImpl* get_nested_tensor_impl_or_null(
+    const at::Tensor& tensor) {
+  if (tensor.is_nested()) {
+    return static_cast<NestedTensorImpl*>(tensor.unsafeGetTensorImpl());
+  }
+  return nullptr;
+}
+
+inline NestedTensorImpl* get_nested_tensor_impl(const at::Tensor& tensor) {
+  TORCH_CHECK(
+      tensor.is_nested(), "get_nested_tensor_impl requires a NestedTensor.");
+  return static_cast<NestedTensorImpl*>(tensor.unsafeGetTensorImpl());
+}
+
+inline bool nested_tensor_impl_is_contiguous(const NestedTensorImpl* nt) {
+  int64_t ntensors = nt->size(0);
+  if (ntensors == 0) {
+    return true;
+  }
+  const Tensor &sizemat = nt->get_nested_sizes(),
+               &stridemat = nt->get_nested_strides();
+  const int64_t* offsets_ptr =
+      nt->get_storage_offsets().const_data_ptr<int64_t>();
+  int64_t orig_dim = sizemat.size(1);
+  // nesting scalars
+  if (orig_dim == 0) {
+    // each scalar must be contiguous
+    // if there is blank memory between underlying scalars
+    for (int64_t i = 0; i < ntensors; i++) {
+      if (offsets_ptr[i] != i) {
+        return false;
+      }
+    }
+  }
+  // nesting tensors
+  else {
+    // if any underlying tensor is non-contiguous
+    const int64_t *sizemat_ptr = sizemat.const_data_ptr<int64_t>(),
+                  *stridemat_ptr = stridemat.const_data_ptr<int64_t>();
+    for (int64_t i = 0; i < ntensors; i++) {
+      if (stridemat_ptr[orig_dim - 1] != 1) {
+        return false;
+      }
+      int64_t product = sizemat_ptr[orig_dim - 1];
+      for (int64_t j = orig_dim - 2; j >= 0; j--) {
+        if (stridemat_ptr[j] != product) {
+          return false;
+        }
+        product *= sizemat_ptr[j];
+      }
+      sizemat_ptr += orig_dim;
+      stridemat_ptr += orig_dim;
+    }
+    // if there is blank memory between underlying tensors
+    if (offsets_ptr[0] != 0) {
+      return false;
+    }
+    sizemat_ptr = sizemat.const_data_ptr<int64_t>();
+    stridemat_ptr = stridemat.const_data_ptr<int64_t>();
+    for (int64_t i = 1; i < ntensors; i++) {
+      if (offsets_ptr[i] !=
+          offsets_ptr[i - 1] + *sizemat_ptr * *stridemat_ptr) {
+        return false;
+      }
+      sizemat_ptr += orig_dim;
+      stridemat_ptr += orig_dim;
+    }
+  }
+  // everything is fine
+  return true;
+}
+
+inline const at::Tensor& get_nested_sizes(const at::Tensor& tensor) {
+  return get_nested_tensor_impl(tensor)->get_nested_sizes();
+}
+
+} // namespace at::native

.venv/lib/python3.11/site-packages/torch/include/ATen/OpMathType.h

@@ -0,0 +1,69 @@
+#pragma once
+
+#include <c10/core/ScalarType.h>
+#include <c10/util/BFloat16.h>
+#include <c10/util/Exception.h>
+#include <c10/util/Float8_e4m3fn.h>
+#include <c10/util/Float8_e4m3fnuz.h>
+#include <c10/util/Float8_e5m2.h>
+#include <c10/util/Float8_e5m2fnuz.h>
+#include <c10/util/Half.h>
+
+namespace at {
+
+// For FP16 or BFloat16 inputs, ops should perform internal math in FP32.
+template <typename scalar_t>
+struct OpMathType {
+  using type = scalar_t;
+};
+template <>
+struct OpMathType<at::Half> {
+  using type = float;
+};
+template <>
+struct OpMathType<at::BFloat16> {
+  using type = float;
+};
+template <>
+struct OpMathType<at::Float8_e5m2> {
+  using type = float;
+};
+template <>
+struct OpMathType<at::Float8_e4m3fn> {
+  using type = float;
+};
+template <>
+struct OpMathType<at::Float8_e5m2fnuz> {
+  using type = float;
+};
+template <>
+struct OpMathType<at::Float8_e4m3fnuz> {
+  using type = float;
+};
+template <>
+struct OpMathType<c10::complex<Half>> {
+  using type = c10::complex<float>;
+};
+
+template <typename T>
+using opmath_type = typename OpMathType<T>::type;
+
+namespace {
+
+inline c10::ScalarType toOpMathType(const c10::ScalarType type) {
+  switch (type) {
+#define DEFINE_CASE(scalar_t, TypeNum) \
+  case ScalarType::TypeNum:            \
+    return CppTypeToScalarType<at::opmath_type<scalar_t>>::value;
+
+    AT_FORALL_SCALAR_TYPES_WITH_COMPLEX(DEFINE_CASE)
+#undef DEFINE_CASE
+
+    default:
+      TORCH_INTERNAL_ASSERT(false, "Unrecognized ScalarType: ", type);
+  }
+}
+
+} // namespace
+
+} // namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/PadNd.h

@@ -0,0 +1,28 @@
+#pragma once
+#include <c10/util/Exception.h>
+#include <c10/util/string_view.h>
+
+namespace at {
+
+enum class padding_mode {
+  reflect,
+  replicate,
+  circular,
+  constant,
+};
+
+static inline c10::string_view padding_mode_string(padding_mode m) {
+  switch (m) {
+    case padding_mode::reflect:
+      return "reflect";
+    case padding_mode::replicate:
+      return "replicate";
+    case padding_mode::circular:
+      return "circular";
+    case padding_mode::constant:
+      return "constant";
+  }
+  TORCH_CHECK(false, "Invalid padding mode (", static_cast<int64_t>(m), ")");
+}
+
+} // namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/Parallel.h

@@ -0,0 +1,158 @@
+#pragma once
+#include <ATen/Config.h>
+#include <c10/macros/Macros.h>
+#include <functional>
+#include <string>
+
+namespace at {
+
+inline int64_t divup(int64_t x, int64_t y) {
+  return (x + y - 1) / y;
+}
+
+// Called during new thread initialization
+TORCH_API void init_num_threads();
+
+// Sets the number of threads to be used in parallel region
+TORCH_API void set_num_threads(int);
+
+// Returns the maximum number of threads that may be used in a parallel region
+TORCH_API int get_num_threads();
+
+// Returns the current thread number (starting from 0)
+// in the current parallel region, or 0 in the sequential region
+TORCH_API int get_thread_num();
+
+// Checks whether the code runs in parallel region
+TORCH_API bool in_parallel_region();
+
+namespace internal {
+
+// Initialise num_threads lazily at first parallel call
+inline void lazy_init_num_threads() {
+  thread_local bool init = false;
+  if (C10_UNLIKELY(!init)) {
+    at::init_num_threads();
+    init = true;
+  }
+}
+
+TORCH_API void set_thread_num(int);
+
+class TORCH_API ThreadIdGuard {
+ public:
+  ThreadIdGuard(int new_id) : old_id_(at::get_thread_num()) {
+    set_thread_num(new_id);
+  }
+
+  ~ThreadIdGuard() {
+    set_thread_num(old_id_);
+  }
+
+ private:
+  int old_id_;
+};
+
+} // namespace internal
+
+/*
+parallel_for
+
+begin: index at which to start applying user function
+
+end: index at which to stop applying user function
+
+grain_size: number of elements per chunk. impacts the degree of parallelization
+
+f: user function applied in parallel to the chunks, signature:
+  void f(int64_t begin, int64_t end)
+
+Warning: parallel_for does NOT copy thread local
+states from the current thread to the worker threads.
+This means for example that Tensor operations CANNOT be used in the
+body of your function, only data pointers.
+*/
+template <class F>
+inline void parallel_for(
+    const int64_t begin,
+    const int64_t end,
+    const int64_t grain_size,
+    const F& f);
+
+/*
+parallel_reduce
+
+begin: index at which to start applying reduction
+
+end: index at which to stop applying reduction
+
+grain_size: number of elements per chunk. impacts number of elements in
+intermediate results tensor and degree of parallelization.
+
+ident: identity for binary combination function sf. sf(ident, x) needs to return
+x.
+
+f: function for reduction over a chunk. f needs to be of signature scalar_t
+f(int64_t partial_begin, int64_t partial_end, scalar_t identifiy)
+
+sf: function to combine two partial results. sf needs to be of signature
+scalar_t sf(scalar_t x, scalar_t y)
+
+For example, you might have a tensor of 10000 entires and want to sum together
+all the elements. Parallel_reduce with a grain_size of 2500 will then allocate
+an intermediate result tensor with 4 elements. Then it will execute the function
+"f" you provide and pass the beginning and end index of these chunks, so
+0-2499, 2500-4999, etc. and the combination identity. It will then write out
+the result from each of these chunks into the intermediate result tensor. After
+that it'll reduce the partial results from each chunk into a single number using
+the combination function sf and the identity ident. For a total summation this
+would be "+" and 0 respectively. This is similar to tbb's approach [1], where
+you need to provide a function to accumulate a subrange, a function to combine
+two partial results and an identity.
+
+Warning: parallel_reduce does NOT copy thread local
+states from the current thread to the worker threads.
+This means for example that Tensor operations CANNOT be used in the
+body of your function, only data pointers.
+
+[1] https://software.intel.com/en-us/node/506154
+*/
+template <class scalar_t, class F, class SF>
+inline scalar_t parallel_reduce(
+    const int64_t begin,
+    const int64_t end,
+    const int64_t grain_size,
+    const scalar_t ident,
+    const F& f,
+    const SF& sf);
+
+// Returns a detailed string describing parallelization settings
+TORCH_API std::string get_parallel_info();
+
+// Sets number of threads used for inter-op parallelism
+TORCH_API void set_num_interop_threads(int);
+
+// Returns the number of threads used for inter-op parallelism
+TORCH_API int get_num_interop_threads();
+
+// Launches inter-op parallel task
+TORCH_API void launch(std::function<void()> func);
+namespace internal {
+void launch_no_thread_state(std::function<void()> fn);
+} // namespace internal
+
+// Launches intra-op parallel task
+TORCH_API void intraop_launch(std::function<void()> func);
+
+// Returns number of intra-op threads used by default
+TORCH_API int intraop_default_num_threads();
+
+} // namespace at
+
+#if AT_PARALLEL_OPENMP
+#include <ATen/ParallelOpenMP.h> // IWYU pragma: keep
+#elif AT_PARALLEL_NATIVE
+#include <ATen/ParallelNative.h> // IWYU pragma: keep
+#endif
+
+#include <ATen/Parallel-inl.h> // IWYU pragma: keep

.venv/lib/python3.11/site-packages/torch/include/ATen/ParallelFuture.h

@@ -0,0 +1,13 @@
+#pragma once
+
+#include <ATen/core/ivalue.h>
+#include <c10/macros/Macros.h>
+#include <functional>
+
+namespace at {
+
+// Launches intra-op parallel task, returns a future
+TORCH_API c10::intrusive_ptr<c10::ivalue::Future> intraop_launch_future(
+    std::function<void()> func);
+
+} // namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/SavedTensorHooks.h

@@ -0,0 +1,66 @@
+#pragma once
+
+#include <c10/core/SafePyObject.h>
+#include <c10/macros/Export.h>
+#include <c10/util/python_stub.h>
+#include <optional>
+#include <stack>
+#include <string>
+
+#include <utility>
+
+namespace at {
+
+namespace impl {
+
+struct TORCH_API SavedTensorDefaultHooksTLS {
+  // PyObject is defined in c10/util/python_stub.h
+  std::stack<std::pair<c10::SafePyObject, c10::SafePyObject>> stack;
+
+  // See NOTE: [Disabling SavedTensorDefaultHooks] for context
+  // NOTE: [disabled_error_message invariant]
+  // disabled_error_message is nullopt IFF Saved Tensor hooks is enabled
+  // We did this for efficiency (so we didn't have to keep a separate bool
+  // around)
+  std::optional<std::string> disabled_error_message;
+
+  // See NOTE: [Deferring tensor pack/unpack hooks until runtime]
+  bool is_tracing = false;
+};
+
+} // namespace impl
+
+struct TORCH_API SavedTensorDefaultHooks {
+  static void push_hooks(
+      c10::SafePyObject pack_hook,
+      c10::SafePyObject unpack_hook);
+  static std::pair<c10::SafePyObject, c10::SafePyObject> pop_hooks();
+  static std::optional<std::pair<c10::SafePyObject, c10::SafePyObject>>
+  get_hooks();
+  static void lazy_initialize();
+
+  static const impl::SavedTensorDefaultHooksTLS& get_tls_state();
+  static void set_tls_state(const impl::SavedTensorDefaultHooksTLS& tls);
+
+  // NOTE: [Disabling SavedTensorDefaultHooks]
+  // A developer of a PyTorch feature may choose to disable SavedTensorDefault
+  // hooks, especially if their feature does not work with it. If they are
+  // disabled, then the following will raise an error:
+  // - Attempting to push_hooks
+  // - calling disable(message) with a non-zero stack (hooks) size
+  static void disable(const std::string& error_message);
+  static void enable();
+  static bool is_enabled();
+  static const std::optional<std::string>& get_disabled_error_message();
+
+  // NOTE: [Deferring tensor pack/unpack hooks until runtime]
+  // To preserve eager semantics of pack/unpack hooks firing only once per saved
+  // variable, Dynamo/AOTAutograd need to defer hook firing until runtime. Using
+  // disable() would loud error at trace time, and pushing a no-op hook would
+  // fail when the traced code is wrapped in a disable_saved_tensors_hooks ctx.
+  // To do so, we disable these hooks during tracing. See
+  // https://github.com/pytorch/pytorch/issues/113263.
+  static bool set_tracing(bool is_tracing);
+};
+
+} // namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/Scalar.h

@@ -0,0 +1,3 @@
+#pragma once
+
+#include <ATen/core/Scalar.h>

.venv/lib/python3.11/site-packages/torch/include/ATen/ScalarOps.h

@@ -0,0 +1,53 @@
+#pragma once
+
+#include <ATen/Tensor.h>
+#include <c10/core/Scalar.h>
+
+#ifndef AT_PER_OPERATOR_HEADERS
+#include <ATen/Functions.h>
+#else
+#include <ATen/ops/scalar_tensor.h>
+#endif
+
+namespace at::detail {
+// When filling a number to 1-element CPU tensor, we want to skip
+// everything but manipulate data ptr directly.
+// Ideally this fast pass should be implemented in TensorIterator,
+// but we also want to skip compute_types which in not avoidable
+// in TensorIterator for now.
+Tensor& scalar_fill(Tensor& self, const Scalar& value);
+TORCH_API Tensor scalar_tensor_static(
+    const Scalar& s,
+    std::optional<ScalarType> dtype_opt,
+    std::optional<Device> device_opt);
+} // namespace at::detail
+
+// This is in the c10 namespace because we use ADL to find the functions in it.
+namespace c10 {
+
+// FIXME: this should be (and was) Scalar::toTensor, but there is currently no
+// way to implement this without going through Derived Types (which are not part
+// of core).
+inline at::Tensor scalar_to_tensor(
+    const Scalar& s,
+    const Device device = at::kCPU) {
+  // This is the fast track we have for CPU scalar tensors.
+  if (device == at::kCPU) {
+    return at::detail::scalar_tensor_static(s, s.type(), at::kCPU);
+  }
+  return at::scalar_tensor(s, at::device(device).dtype(s.type()));
+}
+
+} // namespace c10
+
+namespace at::native {
+
+inline Tensor wrapped_scalar_tensor(
+    const Scalar& scalar,
+    const Device device = at::kCPU) {
+  auto tensor = scalar_to_tensor(scalar, device);
+  tensor.unsafeGetTensorImpl()->set_wrapped_number(true);
+  return tensor;
+}
+
+} // namespace at::native

.venv/lib/python3.11/site-packages/torch/include/ATen/ScalarType.h

@@ -0,0 +1,4 @@
+#pragma once
+#include <ATen/core/ATenGeneral.h> // for BC reasons
+#include <c10/core/Backend.h>
+#include <c10/core/ScalarType.h>

.venv/lib/python3.11/site-packages/torch/include/ATen/SparseCsrTensorImpl.h

@@ -0,0 +1,206 @@
+#pragma once
+
+#include <ATen/Tensor.h>
+#include <c10/core/TensorImpl.h>
+#include <c10/core/impl/TorchDispatchModeTLS.h>
+#include <c10/util/Exception.h>
+namespace at {
+
+// Struct implementing a sparse CSR tensor. It uses three 1-D tensors for
+// denoting the data: `crow_indices_`, `col_indices_` and `values_`.
+// The `crow_indices_` tensor is a integer tensor of shape `(size(0) + 1)`
+// that represents the compressed row indices of the CSR tensor. The
+// `col_indices_` tensor is an integer tensor of shape `(nnz())`
+// that explicitly stores the column indices of each value of the sparse
+// tensor. The `values_` tensor can be of any pytorch-supported data type
+// and has shape `(nnz())`.
+//
+// Since the main advantage of the CSR format over the COO format is speed of
+// computation, care must be taken to facilitate smooth interfacing of
+// these data structures with optimized libraries such as MKL and MAGMA.
+// Since the MKL interface for pytorch currently uses indexing with int32
+// type, it is important to make sure that the `crow_indices` and `col_indices`
+// are of type int32 when calling MKL routines such as SPMM or SPMV.
+//
+// If not calling MKL, it should be alright to use 64 bit integer tensors
+// for indexing.
+struct TORCH_API SparseCsrTensorImpl : public TensorImpl {
+  Tensor crow_indices_;
+  Tensor col_indices_;
+  Tensor values_;
+  Layout layout_;
+
+ public:
+  explicit SparseCsrTensorImpl(
+      at::DispatchKeySet,
+      at::Device device,
+      Layout layout,
+      const caffe2::TypeMeta);
+
+  void resize_(int64_t nnz, IntArrayRef size);
+  void resize_and_clear_(
+      int64_t sparse_dim,
+      int64_t dense_dim,
+      IntArrayRef size);
+  void resize_as_sparse_compressed_tensor_(const Tensor& src);
+  void set_member_tensors(
+      const Tensor& crow_indices,
+      const Tensor& col_indices,
+      const Tensor& values,
+      c10::SymIntArrayRef size);
+  void set_member_tensors(
+      const Tensor& crow_indices,
+      const Tensor& col_indices,
+      const Tensor& values,
+      IntArrayRef size);
+  const Tensor& compressed_indices() const {
+    return crow_indices_;
+  }
+  const Tensor& plain_indices() const {
+    return col_indices_;
+  }
+  const Tensor& values() const {
+    return values_;
+  }
+  int64_t nnz() {
+    return col_indices_.size(-1);
+  }
+
+  inline int64_t batch_dim() const noexcept {
+    return crow_indices_.dim() - 1;
+  }
+
+  inline int64_t sparse_dim() const noexcept {
+    return 2;
+  }
+
+  inline int64_t dense_dim() const noexcept {
+    return values_.dim() - batch_dim() - block_dim() - 1;
+  }
+
+ private:
+  inline int64_t block_dim() const noexcept {
+    return (layout_ == kSparseBsr || layout_ == kSparseBsc ? 2 : 0);
+  }
+
+ protected:
+  IntArrayRef strides_custom() const override;
+  SymIntArrayRef sym_strides_custom() const override;
+  bool is_contiguous_custom(MemoryFormat) const override;
+
+ public:
+  void set_size(int64_t dim, int64_t new_size) override;
+  void set_stride(int64_t dim, int64_t new_stride) override;
+  void set_storage_offset(int64_t storage_offset) override;
+  Layout layout_impl() const override {
+    return layout_;
+  }
+  void set_layout(Layout layout) {
+    switch (layout) {
+      case kSparseCsr:
+      case kSparseCsc:
+      case kSparseBsr:
+      case kSparseBsc:
+        layout_ = layout;
+        break;
+      default:
+        TORCH_CHECK(false, "unsupported layout ", layout);
+    }
+  }
+
+  template <typename VariableVersion>
+  c10::intrusive_ptr<TensorImpl> shallow_copy_and_detach_core(
+      VariableVersion&& version_counter,
+      bool allow_tensor_metadata_change) const {
+    const auto mode_stack_len = c10::impl::TorchDispatchModeTLS::stack_len();
+    c10::impl::PyInterpreter&& interpreter = nullptr;
+    if (mode_stack_len > 0 &&
+        !c10::impl::tls_is_dispatch_key_excluded(DispatchKey::Python)) {
+      const auto& cur_torch_dispatch_mode_state =
+          c10::impl::TorchDispatchModeTLS::get_stack_at(mode_stack_len - 1);
+      interpreter = cur_torch_dispatch_mode_state->pyinterpreter();
+    } else if (
+        key_set_.has(DispatchKey::Python) &&
+        !c10::impl::tls_is_dispatch_key_excluded(DispatchKey::Python)) {
+      interpreter = pyobj_slot_.load_pyobj_interpreter();
+    } else {
+      // otherwise just copy the SparseTensorImpl and not the PyObject.
+      auto impl = c10::make_intrusive<SparseCsrTensorImpl>(
+          key_set(), device(), layout_impl(), dtype());
+      copy_tensor_metadata(
+          /*src_sparse_impl=*/this,
+          /*dest_sparse_impl=*/impl.get(),
+          /*version_counter=*/version_counter,
+          /*allow_tensor_metadata_change=*/allow_tensor_metadata_change);
+      impl->refresh_numel();
+      return impl;
+    }
+    auto r = interpreter->detach(this);
+    r->set_version_counter(std::forward<VariableVersion>(version_counter));
+    r->set_allow_tensor_metadata_change(allow_tensor_metadata_change);
+    return r;
+  }
+
+  /**
+   * Return a TensorImpl that is a shallow-copy of this TensorImpl.
+   *
+   * For usage of `version_counter` and `allow_tensor_metadata_change`,
+   * see NOTE [ TensorImpl Shallow-Copying ].
+   */
+  c10::intrusive_ptr<TensorImpl> shallow_copy_and_detach(
+      const c10::VariableVersion& version_counter,
+      bool allow_tensor_metadata_change) const override {
+    return shallow_copy_and_detach_core(
+        version_counter, allow_tensor_metadata_change);
+  }
+
+  /**
+   * Return a TensorImpl that is a shallow-copy of this TensorImpl.
+   *
+   * For usage of `version_counter` and `allow_tensor_metadata_change`,
+   * see NOTE [ TensorImpl Shallow-Copying ].
+   */
+  c10::intrusive_ptr<TensorImpl> shallow_copy_and_detach(
+      c10::VariableVersion&& version_counter,
+      bool allow_tensor_metadata_change) const override {
+    return shallow_copy_and_detach_core(
+        std::move(version_counter), allow_tensor_metadata_change);
+  }
+
+ private:
+  explicit SparseCsrTensorImpl(
+      at::DispatchKeySet key_set,
+      const caffe2::TypeMeta data_type,
+      at::Tensor crow_indices,
+      at::Tensor col_indices,
+      at::Tensor values,
+      at::Layout layout);
+
+  const char* tensorimpl_type_name() const override;
+
+  /**
+   * Copy the tensor metadata fields (e.g. sizes / strides / storage pointer /
+   * storage_offset) from one TensorImpl to another TensorImpl.
+   *
+   * For usage of `version_counter` and `allow_tensor_metadata_change`, see NOTE
+   * [ TensorImpl Shallow-Copying ].
+   */
+  static void copy_tensor_metadata(
+      const SparseCsrTensorImpl* src_sparse_impl,
+      SparseCsrTensorImpl* dest_sparse_impl,
+      c10::VariableVersion version_counter,
+      bool allow_tensor_metadata_change) {
+    TensorImpl::copy_tensor_metadata(
+        src_sparse_impl,
+        dest_sparse_impl,
+        std::move(version_counter),
+        allow_tensor_metadata_change);
+
+    // Sparse-specific fields
+    dest_sparse_impl->crow_indices_ = src_sparse_impl->compressed_indices();
+    dest_sparse_impl->col_indices_ = src_sparse_impl->plain_indices();
+    dest_sparse_impl->values_ = src_sparse_impl->values();
+    dest_sparse_impl->layout_ = src_sparse_impl->layout_impl();
+  }
+};
+} // namespace at

.venv/lib/python3.11/site-packages/torch/include/ATen/SparseCsrTensorUtils.h

@@ -0,0 +1,441 @@
+#pragma once
+
+#include <ATen/SparseCsrTensorImpl.h>
+#include <ATen/SparseTensorImpl.h>
+#include <ATen/core/Tensor.h>
+
+#ifndef AT_PER_OPERATOR_HEADERS
+#include <ATen/Functions.h>
+#include <ATen/NativeFunctions.h>
+#include <ATen/Operators.h>
+#else
+#include <ATen/ops/_sparse_compressed_tensor_unsafe.h>
+#include <ATen/ops/resize_as_sparse_native.h>
+#endif
+
+#define AT_DISPATCH_ALL_SPARSE_COMPRESSED_LAYOUTS(LAYOUT, NAME, ...) \
+  [&] {                                                              \
+    const auto& the_layout = LAYOUT;                                 \
+    switch (the_layout) {                                            \
+      case kSparseCsr:                                               \
+      case kSparseCsc:                                               \
+      case kSparseBsr:                                               \
+      case kSparseBsc:                                               \
+        return __VA_ARGS__();                                        \
+      default:                                                       \
+        AT_ERROR(                                                    \
+            NAME,                                                    \
+            " expected sparse compressed tensor layout but got ",    \
+            the_layout);                                             \
+    }                                                                \
+  }()
+
+#define AT_DISPATCH_ROW_SPARSE_COMPRESSED_LAYOUTS(                \
+    LAYOUT, NAME, ROW_DIM_ACTION, COLUMN_DIM_ACTION)              \
+  [&]() {                                                         \
+    const auto& the_layout = LAYOUT;                              \
+    switch (the_layout) {                                         \
+      case kSparseCsr:                                            \
+      case kSparseBsr:                                            \
+        return (ROW_DIM_ACTION)();                                \
+      case kSparseCsc:                                            \
+      case kSparseBsc:                                            \
+        return (COLUMN_DIM_ACTION)();                             \
+      default:                                                    \
+        AT_ERROR(                                                 \
+            NAME,                                                 \
+            " expected sparse compressed tensor layout but got ", \
+            the_layout);                                          \
+    }                                                             \
+  }()
+
+#define AT_DISPATCH_PLAIN_SPARSE_COMPRESSED_LAYOUTS(              \
+    LAYOUT, NAME, NO_BLOCK_ACTION, BLOCK_ACTION)                  \
+  [&]() {                                                         \
+    const auto& the_layout = LAYOUT;                              \
+    switch (the_layout) {                                         \
+      case kSparseCsr:                                            \
+      case kSparseCsc:                                            \
+        return (NO_BLOCK_ACTION)();                               \
+      case kSparseBsr:                                            \
+      case kSparseBsc:                                            \
+        return (BLOCK_ACTION)();                                  \
+      default:                                                    \
+        AT_ERROR(                                                 \
+            NAME,                                                 \
+            " expected sparse compressed tensor layout but got ", \
+            the_layout);                                          \
+    }                                                             \
+  }()
+
+#define AT_DISPATCH_SPARSE_ROW_COMPRESSED_LAYOUTS(                    \
+    LAYOUT, NAME, ROW_DIM_ACTION)                                     \
+  [&]() {                                                             \
+    const auto& the_layout = LAYOUT;                                  \
+    switch (the_layout) {                                             \
+      case kSparseCsr:                                                \
+      case kSparseBsr:                                                \
+        return (ROW_DIM_ACTION)();                                    \
+      default:                                                        \
+        AT_ERROR(                                                     \
+            NAME,                                                     \
+            " expected sparse row compressed tensor layout but got ", \
+            the_layout);                                              \
+    }                                                                 \
+  }()
+
+#define AT_DISPATCH_SPARSE_COL_COMPRESSED_LAYOUTS(                       \
+    LAYOUT, NAME, COL_DIM_ACTION)                                        \
+  [&]() {                                                                \
+    const auto& the_layout = LAYOUT;                                     \
+    switch (the_layout) {                                                \
+      case kSparseCsc:                                                   \
+      case kSparseBsc:                                                   \
+        return (COL_DIM_ACTION)();                                       \
+      default:                                                           \
+        AT_ERROR(                                                        \
+            NAME,                                                        \
+            " expected sparse column compressed tensor layout but got ", \
+            the_layout);                                                 \
+    }                                                                    \
+  }()
+
+#define AT_DISPATCH_SPARSE_COMPRESSED_NONBLOCK_LAYOUTS(LAYOUT, NAME, ACTION)  \
+  [&]() {                                                                     \
+    const auto& the_layout = LAYOUT;                                          \
+    switch (the_layout) {                                                     \
+      case kSparseCsr:                                                        \
+      case kSparseCsc:                                                        \
+        return (ACTION)();                                                    \
+      default:                                                                \
+        AT_ERROR(                                                             \
+            NAME,                                                             \
+            " expected sparse compressed (non-block) tensor layout but got ", \
+            the_layout);                                                      \
+    }                                                                         \
+  }()
+
+#define AT_DISPATCH_SPARSE_COMPRESSED_BLOCK_LAYOUTS(LAYOUT, NAME, ACTION) \
+  [&]() {                                                                 \
+    const auto& the_layout = LAYOUT;                                      \
+    switch (the_layout) {                                                 \
+      case kSparseBsr:                                                    \
+      case kSparseBsc:                                                    \
+        return (ACTION)();                                                \
+      default:                                                            \
+        AT_ERROR(                                                         \
+            NAME,                                                         \
+            " expected sparse compressed block tensor layout but got ",   \
+            the_layout);                                                  \
+    }                                                                     \
+  }()
+
+#define AT_DISPATCH_SPARSE_VALUE_TYPES(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                   \
+      TYPE,                                             \
+      NAME,                                             \
+      AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX_AND4(      \
+          kComplexHalf, kHalf, kBool, kBFloat16, __VA_ARGS__))
+
+namespace at::sparse_csr {
+
+// Implements RAII object to manage checking sparse tensor invariants:
+class CheckSparseTensorInvariants {
+  bool old_state;
+
+ public:
+  CheckSparseTensorInvariants(bool state) {
+    old_state = at::globalContext().checkSparseTensorInvariants();
+    at::globalContext().setCheckSparseTensorInvariants(state);
+  }
+
+  ~CheckSparseTensorInvariants() {
+    at::globalContext().setCheckSparseTensorInvariants(old_state);
+  }
+};
+
+using SparseCsrTensor = Tensor;
+
+inline bool is_sparse_compressed(const Layout& layout) {
+  switch (layout) {
+    case kSparseCsr:
+    case kSparseCsc:
+    case kSparseBsr:
+    case kSparseBsc:
+      return true;
+    default:;
+  }
+  return false;
+}
+
+inline bool is_sparse_compressed(const Tensor& self) {
+  return is_sparse_compressed(self.layout());
+}
+
+inline SparseCsrTensorImpl* get_sparse_csr_impl(const SparseCsrTensor& self) {
+  AT_DISPATCH_ALL_SPARSE_COMPRESSED_LAYOUTS(
+      self.layout(), "get_sparse_csr_impl", [&] {});
+  return static_cast<SparseCsrTensorImpl*>(self.unsafeGetTensorImpl());
+}
+
+inline std::string layoutToString(
+    Layout layout,
+    bool upper = false,
+    bool lower = false) {
+  switch (layout) {
+    case kSparseCsr:
+      return (upper ? "CSR" : (lower ? "csr" : "Csr"));
+    case kSparseCsc:
+      return (upper ? "CSC" : (lower ? "csc" : "Csc"));
+    case kSparseBsr:
+      return (upper ? "BSR" : (lower ? "bsr" : "Bsr"));
+    case kSparseBsc:
+      return (upper ? "BSC" : (lower ? "bsc" : "Bsc"));
+    default:
+      TORCH_CHECK(false, "Not a sparse compressed layout:", layout);
+      return "";
+  }
+}
+
+inline bool isCompressedRow(Layout layout) {
+  return AT_DISPATCH_ROW_SPARSE_COMPRESSED_LAYOUTS(
+      layout, "isCompressedRow", [&] { return true; }, [&] { return false; });
+}
+
+inline bool isCompressedColumn(Layout layout) {
+  return AT_DISPATCH_ROW_SPARSE_COMPRESSED_LAYOUTS(
+      layout,
+      "isCompressedColumn",
+      [&] { return false; },
+      [&] { return true; });
+}
+
+inline std::string compressedIndicesName(Layout layout) {
+  return AT_DISPATCH_ROW_SPARSE_COMPRESSED_LAYOUTS(
+      layout,
+      "compressedIndicesName",
+      [&] { return "crow_indices"; },
+      [&] { return "ccol_indices"; });
+}
+
+inline std::string plainIndicesName(Layout layout) {
+  return AT_DISPATCH_ROW_SPARSE_COMPRESSED_LAYOUTS(
+      layout,
+      "plainIndicesName",
+      [&] { return "col_indices"; },
+      [&] { return "row_indices"; });
+}
+
+inline std::string compressedDimName(Layout layout) {
+  switch (layout) {
+    case kSparseCsr:
+      return "row";
+    case kSparseCsc:
+      return "column";
+    case kSparseBsr:
+      return "row block";
+    case kSparseBsc:
+      return "column block";
+    default:
+      TORCH_CHECK(false, "Not a sparse compressed layout:", layout);
+      return "";
+  }
+}
+
+inline std::string plainDimName(Layout layout) {
+  switch (layout) {
+    case kSparseCsr:
+      return "column";
+    case kSparseCsc:
+      return "row";
+    case kSparseBsr:
+      return "column block";
+    case kSparseBsc:
+      return "row block";
+    default:
+      TORCH_CHECK(false, "Not a sparse compressed layout:", layout);
+      return "";
+  }
+}
+
+inline size_t rowDimension(Layout layout, IntArrayRef size) {
+  return size.size() - (isCompressedRow(layout) ? 2 : 1);
+}
+
+inline size_t columnDimension(Layout layout, IntArrayRef size) {
+  return size.size() - (isCompressedColumn(layout) ? 2 : 1);
+}
+
+inline size_t compressedDimension(
+    Layout layout,
+    IntArrayRef size,
+    size_t dense_ndim = 0) {
+  return size.size() - dense_ndim - (isCompressedRow(layout) ? 2 : 1);
+}
+
+inline size_t plainDimension(
+    Layout layout,
+    IntArrayRef size,
+    size_t dense_ndim = 0) {
+  return size.size() - dense_ndim - (isCompressedRow(layout) ? 1 : 2);
+}
+
+inline int64_t numBatchDimensions(Tensor const& self) {
+  return AT_DISPATCH_ROW_SPARSE_COMPRESSED_LAYOUTS(
+      self.layout(),
+      "numBatchDimensions",
+      [&self] { return self.crow_indices().dim() - 1; },
+      [&self] { return self.ccol_indices().dim() - 1; });
+}
+
+inline std::pair<Tensor, Tensor> getCompressedPlainIndices(Tensor const& self) {
+  return AT_DISPATCH_ROW_SPARSE_COMPRESSED_LAYOUTS(
+      self.layout(),
+      "getCompressedPlainIndices",
+      [&self] {
+        return std::make_pair(self.crow_indices(), self.col_indices());
+      },
+      [&self] {
+        return std::make_pair(self.ccol_indices(), self.row_indices());
+      });
+}
+
+inline ScalarType getIndexDtype(Tensor const& self) {
+  switch (self.layout()) {
+    case kSparseCsr:
+    case kSparseBsr:
+      return self.crow_indices().scalar_type();
+    case kSparseCsc:
+    case kSparseBsc:
+      return self.ccol_indices().scalar_type();
+    case kSparse:
+      return self._indices().scalar_type();
+    default:
+      return ScalarType::Long;
+  }
+}
+
+inline Layout flip_compressed_layout(Layout layout) {
+  switch (layout) {
+    case kSparseCsr:
+      return kSparseCsc;
+    case kSparseCsc:
+      return kSparseCsr;
+    case kSparseBsr:
+      return kSparseBsc;
+    case kSparseBsc:
+      return kSparseBsr;
+    default:
+      TORCH_CHECK(false, "Not a sparse compressed layout:", layout);
+      return kSparseCsr;
+  }
+}
+
+inline DimVector getBlockSize(Tensor const& self) {
+  int64_t n_batch = numBatchDimensions(self);
+  return at::DimVector(self.values().sizes().slice(n_batch + 1, 2));
+}
+
+inline at::OptionalArray<at::SymInt> getSymIntBlockSize(Tensor const& self) {
+  if (self.layout() == at::kSparseBsr || self.layout() == at::kSparseBsc) {
+    int64_t n_batch = numBatchDimensions(self);
+    return self.values().sym_sizes().slice(n_batch + 1, 2).vec();
+  } else {
+    return {};
+  }
+}
+
+template <typename binary_op_t, typename binary_op_out_t>
+inline bool only_sparse_compressed_binary_op_trivial_cases(
+    const Tensor& self,
+    const Tensor& other,
+    const Scalar& alpha,
+    Tensor& out,
+    const binary_op_t& binary_op,
+    const binary_op_out_t& binary_op_out) {
+  // Only sparse compressed! Just like the name says :)
+  TORCH_INTERNAL_ASSERT(at::sparse_csr::is_sparse_compressed(self));
+  TORCH_INTERNAL_ASSERT(at::sparse_csr::is_sparse_compressed(other));
+  TORCH_INTERNAL_ASSERT(at::sparse_csr::is_sparse_compressed(out));
+
+  // Bypass BLAS if there are matches in (self, other, out)
+  if (self.is_same(out) && self.is_same(other)) {
+    binary_op_out(self.values(), other.values(), alpha);
+    return true;
+  }
+  if (self.is_same(other)) {
+    auto [compressed_indices, plain_indices] =
+        at::sparse_csr::getCompressedPlainIndices(self);
+    static_cast<SparseCsrTensorImpl*>(out.unsafeGetTensorImpl())
+        ->set_member_tensors(
+            compressed_indices,
+            plain_indices,
+            binary_op(self.values(), other.values(), alpha),
+            self.sizes());
+    return true;
+  }
+  return false;
+}
+
+inline bool only_sparse_compressed_add_trivial_cases(
+    const Tensor& self,
+    const Tensor& other,
+    const Scalar& alpha,
+    Tensor& out) {
+  return only_sparse_compressed_binary_op_trivial_cases(
+      self,
+      other,
+      alpha,
+      out,
+      [](const Tensor& v1, const Tensor& v2, const Scalar& alpha) {
+        return v1.add(v2, alpha);
+      },
+      [](const Tensor& v1, const Tensor& v2, const Scalar& alpha) {
+        return v1.add_(v2, alpha);
+      });
+}
+
+inline Tensor to_type(const Tensor& input, ScalarType dtype) {
+  auto [compressed_indices, plain_indices] =
+      at::sparse_csr::getCompressedPlainIndices(input);
+  return at::_sparse_compressed_tensor_unsafe(
+      compressed_indices,
+      plain_indices,
+      std::move(input.values()).to(dtype),
+      input.sizes(),
+      dtype,
+      input.layout(),
+      input.device(),
+      input.options().pinned_memory_opt());
+}
+
+template <typename acc_t, typename scalar_t>
+inline std::tuple<Tensor, Tensor> create_acc_buffer(
+    TensorOptions option,
+    ScalarType type,
+    int64_t nnz = -1) {
+  Tensor new_values, new_values_acc;
+  constexpr bool need_acc = !std::is_same_v<scalar_t, acc_t>;
+  bool is_integral = at::isIntegralType(type, /*includeBool=*/true);
+  if constexpr (need_acc) {
+    auto acc_dtype = CppTypeToScalarType<acc_t>::value;
+    new_values_acc = at::empty({}, option.dtype(acc_dtype));
+    new_values = is_integral ? new_values_acc : at::empty({}, option);
+  } else {
+    new_values = new_values_acc = at::empty({}, option);
+  }
+  if (nnz != -1) {
+    return std::make_tuple(
+        new_values.resize_(nnz), new_values_acc.resize_(nnz));
+  } else {
+    return std::make_tuple(new_values, new_values_acc);
+  }
+}
+
+inline void copy_from_acc_buffer(Tensor& new_values, Tensor& new_values_acc) {
+  if (!new_values_acc.is_same(new_values)) {
+    new_values.copy_(new_values_acc);
+  }
+}
+
+} // namespace at::sparse_csr

.venv/lib/python3.11/site-packages/torch/include/ATen/Storage.h

@@ -0,0 +1,2 @@
+#pragma once
+#include <c10/core/Storage.h>

.venv/lib/python3.11/site-packages/torch/include/ATen/Tensor.h

@@ -0,0 +1,3 @@
+#pragma once
+
+#include <ATen/core/Tensor.h>