Add files using upload-large-folder tool

phivenv/Lib/site-packages/torch/include/oneapi/dnnl/dnnl_graph_sycl.h

@@ -0,0 +1,99 @@
+/*******************************************************************************
+* Copyright 2020-2024 Intel Corporation
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*******************************************************************************/
+
+#ifndef ONEAPI_DNNL_DNNL_GRAPH_SYCL_H
+#define ONEAPI_DNNL_DNNL_GRAPH_SYCL_H
+
+#include "oneapi/dnnl/dnnl_graph.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/// @addtogroup dnnl_api
+/// @{
+
+/// @addtogroup dnnl_graph_api
+/// @{
+
+/// @addtogroup dnnl_graph_api_interop
+/// @{
+
+/// @addtogroup dnnl_graph_api_sycl_interop
+/// @{
+
+/// Allocation call-back function interface for SYCL. SYCL allocator should be
+/// used for SYCL runtime and host allocator should be used for non-SYCL. The
+/// call-back should return a USM device memory pointer.
+typedef void *(*dnnl_graph_sycl_allocate_f)(
+        size_t size, size_t alignment, const void *dev, const void *context);
+
+/// Deallocation call-back function interface for SYCL. SYCL allocator should be
+/// used for SYCL runtime and host allocator should be used for non-SYCL. The
+/// call-back should deallocate a USM device memory returned by
+/// #dnnl_graph_sycl_allocate_f.
+typedef void (*dnnl_graph_sycl_deallocate_f)(
+        void *buf, const void *dev, const void *context, void *event);
+
+/// Creates an allocator with the given allocation and deallocation call-back
+/// function pointers.
+///
+/// @param allocator Output allocator
+/// @param sycl_malloc A pointer to SYCL malloc function
+/// @param sycl_free A pointer to SYCL free function
+/// @returns #dnnl_success on success and a status describing the
+///     error otherwise.
+dnnl_status_t DNNL_API dnnl_graph_sycl_interop_allocator_create(
+        dnnl_graph_allocator_t *allocator,
+        dnnl_graph_sycl_allocate_f sycl_malloc,
+        dnnl_graph_sycl_deallocate_f sycl_free);
+
+/// This API is a supplement for existing onednn engine API.
+dnnl_status_t DNNL_API dnnl_graph_sycl_interop_make_engine_with_allocator(
+        dnnl_engine_t *engine, const void *device, const void *context,
+        const_dnnl_graph_allocator_t alloc);
+
+/// Execute a compiled partition with sycl runtime.
+///
+/// @param compiled_partition The handle of target compiled_partition.
+/// @param stream The stream used for execution
+/// @param num_inputs The number of input tensors
+/// @param inputs A list of input tensors
+/// @param num_outputs The number of output tensors
+/// @param outputs A non-empty list of output tensors
+/// @param deps Optional handle of list with `sycl::event` dependencies.
+/// @param sycl_event The handle of sycl event.
+/// @returns #dnnl_success on success and a status describing the
+///     error otherwise.
+dnnl_status_t DNNL_API dnnl_graph_sycl_interop_compiled_partition_execute(
+        const_dnnl_graph_compiled_partition_t compiled_partition,
+        dnnl_stream_t stream, size_t num_inputs,
+        const_dnnl_graph_tensor_t *inputs, size_t num_outputs,
+        const_dnnl_graph_tensor_t *outputs, const void *deps, void *sycl_event);
+
+/// @} dnnl_graph_api_sycl_interop
+
+/// @} dnnl_graph_api_interop
+
+/// @} dnnl_graph_api
+
+/// @} dnnl_api
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif

phivenv/Lib/site-packages/torch/include/oneapi/dnnl/dnnl_graph_sycl.hpp

@@ -0,0 +1,131 @@
+/*******************************************************************************
+* Copyright 2020-2025 Intel Corporation
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*******************************************************************************/
+
+/// @file
+/// Graph SYCL interop API
+
+#ifndef ONEAPI_DNNL_DNNL_GRAPH_SYCL_HPP
+#define ONEAPI_DNNL_DNNL_GRAPH_SYCL_HPP
+
+/// @cond DO_NOT_DOCUMENT_THIS
+#include <vector>
+
+#if __has_include(<sycl/sycl.hpp>)
+#include <sycl/sycl.hpp>
+#else
+#error "Unsupported compiler"
+#endif
+
+#include "oneapi/dnnl/dnnl_graph.hpp"
+#include "oneapi/dnnl/dnnl_graph_sycl.h"
+/// @endcond
+
+/// @addtogroup dnnl_api
+/// @{
+
+namespace dnnl {
+
+/// @addtogroup dnnl_graph_api
+/// @{
+
+namespace graph {
+
+/// @addtogroup dnnl_graph_api_interop Runtime interoperability API
+/// API extensions to interact with the underlying run-time.
+/// @{
+
+/// @addtogroup dnnl_graph_api_sycl_interop SYCL interoperability API
+/// API extensions to interact with the underlying SYCL run-time.
+/// @{
+
+/// SYCL interoperability namespace
+namespace sycl_interop {
+
+/// Constructs an allocator from SYCL malloc and free function pointer. SYCL
+/// allocator  should be used for SYCL runtime and host allocator should be used
+/// for non-SYCL. Currently, only device USM allocator is supported.
+///
+/// @param sycl_malloc The pointer to SYCL malloc function
+/// @param sycl_free The pointer to SYCL free function
+/// @returns Created allocator
+inline allocator make_allocator(dnnl_graph_sycl_allocate_f sycl_malloc,
+        dnnl_graph_sycl_deallocate_f sycl_free) {
+    dnnl_graph_allocator_t c_allocator = nullptr;
+    error::wrap_c_api(dnnl_graph_sycl_interop_allocator_create(
+                              &c_allocator, sycl_malloc, sycl_free),
+            "could not create allocator for sycl device");
+    return allocator(c_allocator);
+}
+
+inline engine make_engine_with_allocator(const sycl::device &adevice,
+        const sycl::context &acontext, const allocator &alloc) {
+    dnnl_engine_t c_engine;
+    error::wrap_c_api(
+            dnnl_graph_sycl_interop_make_engine_with_allocator(&c_engine,
+                    static_cast<const void *>(&adevice),
+                    static_cast<const void *>(&acontext), alloc.get()),
+            "could not make an engine with allocator");
+    return engine(c_engine);
+}
+
+/// Executes a compiled partition in a specified stream and returns a SYCL
+/// event.
+///
+/// @param c_partition Compiled partition to execute.
+/// @param astream Stream object to run over
+/// @param inputs Arguments map.
+/// @param outputs Arguments map.
+/// @param deps Optional vector with `sycl::event` dependencies.
+/// @returns Output event.
+inline sycl::event execute(compiled_partition &c_partition, stream &astream,
+        const std::vector<tensor> &inputs, std::vector<tensor> &outputs,
+        const std::vector<sycl::event> &deps = {}) {
+    std::vector<const_dnnl_graph_tensor_t> c_inputs;
+    c_inputs.reserve(inputs.size());
+    for (auto &in : inputs) {
+        c_inputs.push_back(in.get());
+    }
+    std::vector<const_dnnl_graph_tensor_t> c_outputs;
+    c_outputs.reserve(outputs.size());
+    for (auto &out : outputs) {
+        c_outputs.push_back(out.get());
+    }
+
+    sycl::event sycl_event;
+    error::wrap_c_api(dnnl_graph_sycl_interop_compiled_partition_execute(
+                              c_partition.get(), astream.get(), c_inputs.size(),
+                              c_inputs.data(), c_outputs.size(),
+                              c_outputs.data(), &deps, &sycl_event),
+            "could not execute the compiled_partition on a specified sycl "
+            "stream");
+    return sycl_event;
+}
+
+} // namespace sycl_interop
+
+/// @} dnnl_graph_api_sycl_interop
+
+/// @} dnnl_graph_api_interop
+
+} // namespace graph
+
+/// @} dnnl_graph_api
+
+} // namespace dnnl
+
+/// @} dnnl_api
+
+#endif

phivenv/Lib/site-packages/torch/include/oneapi/dnnl/dnnl_graph_types.h

@@ -0,0 +1,475 @@
+/*******************************************************************************
+ * Copyright 2020-2025 Intel Corporation
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ *******************************************************************************/
+
+/// @file
+/// C API definitions
+
+#ifndef ONEAPI_DNNL_DNNL_GRAPH_TYPES_H
+#define ONEAPI_DNNL_DNNL_GRAPH_TYPES_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/// @cond DO_NOT_DOCUMENT_THIS
+#include <limits.h>
+#include <stddef.h>
+
+#include "oneapi/dnnl/dnnl_common_types.h"
+/// @endcond
+
+/// @addtogroup dnnl_api
+/// @{
+
+/// @addtogroup dnnl_graph_api
+/// @{
+
+/// @addtogroup dnnl_graph_api_logical_tensor
+/// @{
+
+/// A wildcard value for number of dimensions which is unknown at a tensor or
+/// operation creation time.
+#define DNNL_GRAPH_UNKNOWN_NDIMS -1
+
+/// A wildcard value for dimensions that are unknown at a tensor or operation
+/// creation time.
+#define DNNL_GRAPH_UNKNOWN_DIM INT64_MIN
+
+/// Layout type specification
+typedef enum {
+    /// Undefined layout type
+    dnnl_graph_layout_type_undef = 0,
+    /// Any means to let the library to decide the layout for a tensor during
+    /// partition compilation.
+    dnnl_graph_layout_type_any = 1,
+    /// Strided means that the layout of a tensor is determined by the strides
+    /// field in the logical tensor.
+    dnnl_graph_layout_type_strided = 2,
+    /// Opaque means that the layout of a tensor is the library specific.
+    /// Usually, an opaque layout is generated by a partition which is compiled
+    /// with layout type any.
+    dnnl_graph_layout_type_opaque = 3,
+} dnnl_graph_layout_type_t;
+
+/// Logical tensor property
+typedef enum {
+    /// Undefined tensor property
+    dnnl_graph_tensor_property_undef = 0,
+    /// Variable means the tensor may be changed during computation or between
+    /// different iterations.
+    dnnl_graph_tensor_property_variable = 1,
+    /// Constant means the tensor will keep unchanged during computation and
+    /// between different iterations. It's useful for the library to apply
+    /// optimizations for constant tensors or cache constant tensors inside the
+    /// library. For example, constant weight tensors in inference scenarios.
+    dnnl_graph_tensor_property_constant = 2,
+} dnnl_graph_tensor_property_t;
+
+/// Logical tensor. It is based on an ID, a number of dimensions, dimensions
+/// themselves, element data type, tensor property and tensor memory layout.
+typedef struct {
+    /// Unique id of each logical tensor. The library uses logical tensor IDs to
+    /// build up the connections between operations if the output of one
+    /// operation has the same ID as the input of another operation.
+    size_t id;
+
+    /// Number of dimensions. -1 means unknown (DNNL_GRAPH_UNKNOWN_NDIMS). 0 is
+    /// used to define scalar tensor.
+    int ndims;
+
+    /// Size of each dimension. #DNNL_GRAPH_UNKNOWN_DIM means the size of that
+    /// dimension is unknown. 0 is used to define zero-dimension tensor. The
+    /// library supports to deduce output shapes according to input shapes
+    /// during compilation. Unlike memory descriptor in oneDNN primitive API,
+    /// the order of dimensions is not defined in logical tensor. It is defined
+    /// by the operations which respect the order through the attributes
+    /// #dnnl_graph_op_attr_data_format or #dnnl_graph_op_attr_weights_format.
+    /// For example, for a Convolution with `data_format=NXC`, it means the
+    /// first element of dims of activation tensor is mini-batch size, the last
+    /// effective element of dims is channel size, and other elements between
+    /// them are spatial dimensions.
+    dnnl_dims_t dims;
+
+    /// Data type of the tensor elements.
+    dnnl_data_type_t data_type;
+
+    /// Property type of the tensor.
+    dnnl_graph_tensor_property_t property;
+
+    /// Layout type of the tensor.
+    dnnl_graph_layout_type_t layout_type;
+    union {
+        /// The field is valid when `layout_type` is
+        /// #dnnl_graph_layout_type_strided. #DNNL_GRAPH_UNKNOWN_DIM means the
+        /// stride of the dimension is unknown. The library currently doesn't
+        /// support other negative stride values.
+        dnnl_dims_t strides;
+
+        /// The field is valid when `layout_type` is
+        /// #dnnl_graph_layout_type_opaque. An opaque layout ID is usually
+        /// generated by a partition which is compiled with layout type any.
+        size_t layout_id;
+    } layout;
+} dnnl_graph_logical_tensor_t;
+
+/// @} dnnl_graph_api_logical_tensor
+
+/// @addtogroup dnnl_graph_api_partition
+/// @{
+
+/// Policy specifications for partitioning
+typedef enum {
+    /// Fusion policy returns partitions with typical post-op fusions, eg.
+    /// Convolution + ReLU or other element-wise operations or a chian of
+    /// post-ops.
+    dnnl_graph_partition_policy_fusion = 1,
+    /// Debug policy doesn't not apply any fusions. It returns partitions with
+    /// single operation in each partition. The policy is useful when users
+    /// notice any bug or correctness issue in fusion policy.
+    dnnl_graph_partition_policy_debug = 2,
+} dnnl_graph_partition_policy_t;
+
+/// An opaque structure to describe a partition.
+struct dnnl_graph_partition;
+
+/// A partition handle.
+typedef struct dnnl_graph_partition *dnnl_graph_partition_t;
+
+/// A constant partition handle.
+typedef const struct dnnl_graph_partition *const_dnnl_graph_partition_t;
+
+/// @} dnnl_graph_api_partition
+
+/// @addtogroup dnnl_graph_api_graph
+/// @{
+
+/// An opaque structure to describe a graph.
+struct dnnl_graph_graph;
+
+/// A graph handle.
+typedef struct dnnl_graph_graph *dnnl_graph_graph_t;
+
+/// A constant graph handle.
+typedef const struct dnnl_graph_graph *const_dnnl_graph_graph_t;
+
+/// @} dnnl_graph_api_graph
+
+/// @addtogroup dnnl_graph_api_op
+/// @{
+
+/// Kinds of operations
+typedef enum {
+    dnnl_graph_op_abs,
+    dnnl_graph_op_abs_backward,
+    dnnl_graph_op_add,
+    dnnl_graph_op_avg_pool,
+    dnnl_graph_op_avg_pool_backward,
+    dnnl_graph_op_batch_norm_backward,
+    dnnl_graph_op_batch_norm_forward_training,
+    dnnl_graph_op_batch_norm_inference,
+    dnnl_graph_op_bias_add,
+    dnnl_graph_op_bias_add_backward,
+    dnnl_graph_op_clamp,
+    dnnl_graph_op_clamp_backward,
+    dnnl_graph_op_concat,
+    dnnl_graph_op_convolution,
+    dnnl_graph_op_convolution_backward_data,
+    dnnl_graph_op_convolution_backward_weights,
+    dnnl_graph_op_conv_transpose,
+    dnnl_graph_op_conv_transpose_backward_data,
+    dnnl_graph_op_conv_transpose_backward_weights,
+    dnnl_graph_op_dequantize,
+    dnnl_graph_op_divide,
+    dnnl_graph_op_dynamic_dequantize,
+    dnnl_graph_op_dynamic_quantize,
+    dnnl_graph_op_elu,
+    dnnl_graph_op_elu_backward,
+    dnnl_graph_op_end,
+    dnnl_graph_op_exp,
+    dnnl_graph_op_gelu,
+    dnnl_graph_op_gelu_backward,
+    dnnl_graph_op_hard_swish,
+    dnnl_graph_op_hard_swish_backward,
+    dnnl_graph_op_interpolate,
+    dnnl_graph_op_interpolate_backward,
+    dnnl_graph_op_layer_norm,
+    dnnl_graph_op_layer_norm_backward,
+    dnnl_graph_op_leaky_relu,
+    dnnl_graph_op_log,
+    dnnl_graph_op_log_softmax,
+    dnnl_graph_op_log_softmax_backward,
+    dnnl_graph_op_matmul,
+    dnnl_graph_op_maximum,
+    dnnl_graph_op_max_pool,
+    dnnl_graph_op_max_pool_backward,
+    dnnl_graph_op_minimum,
+    dnnl_graph_op_mish,
+    dnnl_graph_op_mish_backward,
+    dnnl_graph_op_multiply,
+    dnnl_graph_op_prelu,
+    dnnl_graph_op_prelu_backward,
+    dnnl_graph_op_quantize,
+    dnnl_graph_op_reciprocal,
+    dnnl_graph_op_reduce_l1,
+    dnnl_graph_op_reduce_l2,
+    dnnl_graph_op_reduce_max,
+    dnnl_graph_op_reduce_mean,
+    dnnl_graph_op_reduce_min,
+    dnnl_graph_op_reduce_prod,
+    dnnl_graph_op_reduce_sum,
+    dnnl_graph_op_relu,
+    dnnl_graph_op_relu_backward,
+    dnnl_graph_op_reorder,
+    dnnl_graph_op_round,
+    dnnl_graph_op_sigmoid,
+    dnnl_graph_op_sigmoid_backward,
+    dnnl_graph_op_softmax,
+    dnnl_graph_op_softmax_backward,
+    dnnl_graph_op_softplus,
+    dnnl_graph_op_softplus_backward,
+    dnnl_graph_op_sqrt,
+    dnnl_graph_op_sqrt_backward,
+    dnnl_graph_op_square,
+    dnnl_graph_op_squared_difference,
+    dnnl_graph_op_static_reshape,
+    dnnl_graph_op_static_transpose,
+    dnnl_graph_op_subtract,
+    dnnl_graph_op_tanh,
+    dnnl_graph_op_tanh_backward,
+    dnnl_graph_op_type_cast,
+    dnnl_graph_op_wildcard,
+    dnnl_graph_op_hard_sigmoid,
+    dnnl_graph_op_hard_sigmoid_backward,
+    dnnl_graph_op_select,
+    dnnl_graph_op_pow,
+    dnnl_graph_op_group_norm,
+    dnnl_graph_op_gen_index,
+    dnnl_graph_op_greater_equal,
+    dnnl_graph_op_last_symbol,
+} dnnl_graph_op_kind_t;
+
+/// Attributes of operations
+typedef enum {
+    /// Undefined op attribute.
+    dnnl_graph_op_attr_undef = 0,
+
+    // float32 attributes. The value of these attributes can be any single
+    // float32 number.
+
+    /// Specifies an alpha attribute to an op.
+    dnnl_graph_op_attr_alpha = 0x1,
+    /// Specifies an beta attribute to an op.
+    dnnl_graph_op_attr_beta,
+    /// Specifies an epsilon attribute to an op.
+    dnnl_graph_op_attr_epsilon,
+    /// Specifies a max attribute to an op.
+    dnnl_graph_op_attr_max,
+    ///Specifies a min attribute to an op.
+    dnnl_graph_op_attr_min,
+    /// Specifies a momentum attribute to an op.
+    dnnl_graph_op_attr_momentum,
+
+    // float32 vector attributes. The value of these attributes can be a vector
+    // of float32 numbers.
+
+    /// Specifies a scales attribute to an op.
+    dnnl_graph_op_attr_scales = 0x20,
+
+    // int64_t attributes. The value of these attributes can be any single int64
+    // number.
+
+    /// Specifies an axis attribute to an op.
+    dnnl_graph_op_attr_axis = 0x30,
+    /// Specifies a begin_norm_axis attribute to an op.
+    dnnl_graph_op_attr_begin_norm_axis,
+    /// Specifies a groups attribute to an op.
+    dnnl_graph_op_attr_groups,
+
+    // int64_t vector attributes. The value of these attributes can be a vector
+    // of int64 numbers.
+
+    /// Specifies an axes attribute to an op.
+    dnnl_graph_op_attr_axes = 0x40,
+    /// Specifies a dilations attribute to an op.
+    dnnl_graph_op_attr_dilations,
+    /// Specifies an dst_shape attribute to an op.
+    dnnl_graph_op_attr_dst_shape,
+    /// Specifies a kernel attribute to an op.
+    dnnl_graph_op_attr_kernel,
+    /// Specifies an order attribute to an op.
+    dnnl_graph_op_attr_order,
+    /// Specifies an output_padding attribute to an op.
+    dnnl_graph_op_attr_output_padding,
+    /// Specifies a pads_begin attribute to an op.
+    dnnl_graph_op_attr_pads_begin,
+    /// Specifies a pads_end attribute to an op.
+    dnnl_graph_op_attr_pads_end,
+    /// Specifies a shape attribute to an op.
+    dnnl_graph_op_attr_shape,
+    /// Specifies a sizes attribute to an op.
+    dnnl_graph_op_attr_sizes,
+    /// Specifies a input_shape attribute to an op.
+    dnnl_graph_op_attr_src_shape,
+    /// Specifies a strides attribute to an op.
+    dnnl_graph_op_attr_strides,
+    /// Specifies a weight_shape attribute to an op.
+    dnnl_graph_op_attr_weights_shape,
+    /// Specifies a zps attribute to an op.
+    dnnl_graph_op_attr_zps,
+    /// Specifies a group shape attribute to an op.
+    dnnl_graph_op_attr_group_shape,
+
+    // bool attributes. The value of these attributes can be any single bool
+    // value.
+
+    /// Specifies an exclude_pad attribute to an op.
+    dnnl_graph_op_attr_exclude_pad = 0x60,
+    /// Specifies a keep_dims attribute to an op.
+    dnnl_graph_op_attr_keep_dims,
+    /// Specifies a keep_stats attribute to an op.
+    dnnl_graph_op_attr_keep_stats,
+    /// Specifies a per_channel_broadcast attribute to an op.
+    dnnl_graph_op_attr_per_channel_broadcast,
+    /// Specifies a special_zero attribute to an op.
+    dnnl_graph_op_attr_special_zero,
+    /// Specifies a transpose_a attribute to an op.
+    dnnl_graph_op_attr_transpose_a,
+    /// Specifies a transpose_b attribute to an op.
+    dnnl_graph_op_attr_transpose_b,
+    /// Specifies an use_affine attribute to an op.
+    dnnl_graph_op_attr_use_affine,
+    /// Specifies an use_dst attribute to an op.
+    dnnl_graph_op_attr_use_dst,
+
+    // string attributes. The value of these attributes can be a string.
+
+    /// Specifies an auto_broadcast attribute to an op. The value can be "none"
+    /// or "numpy".
+    dnnl_graph_op_attr_auto_broadcast = 0x80,
+    /// Specifies an auto_pad attribute to an op. The value can be "none",
+    /// "same_upper", "same_lower", or "valid".
+    dnnl_graph_op_attr_auto_pad,
+    /// Specifies an coordinate_transformation_mode attribute to an op. The
+    /// value can be "half_pixel" or "align_corners". The attribute is defined
+    /// for Interpolate operations.
+    dnnl_graph_op_attr_coordinate_transformation_mode,
+    /// Specifies a data_format of an op. The value can be "NCX" or "NXC".
+    dnnl_graph_op_attr_data_format,
+    /// Specifies a mode attribute of an op. The value can be "nearest",
+    /// "linear", "bilinear", or "trilinear". The attribute is defined for
+    /// Interpolate operations.
+    dnnl_graph_op_attr_mode,
+    /// Specifies a qtype attribute to an op. The value can be "per_channel" or
+    /// "per_tensor". The attribute is defined for quantization operations.
+    dnnl_graph_op_attr_qtype,
+    /// Specifies a rounding_type attribute to an op. The value can be "ceil" or
+    /// "floor".
+    dnnl_graph_op_attr_rounding_type,
+    /// Specifies a weights_format of an op. The value can be "OIX", "XIO",
+    /// "IOX", or "XOI". Different operations may support different values.
+    dnnl_graph_op_attr_weights_format,
+
+    /// Specifies the end of all above exteral attributes for check.
+    dnnl_graph_op_attr_end = 0xFF,
+} dnnl_graph_op_attr_t;
+
+/// An opaque structure to describe an operation.
+struct dnnl_graph_op;
+
+/// An operation handle.
+typedef struct dnnl_graph_op *dnnl_graph_op_t;
+
+/// A constant operation handle.
+typedef const struct dnnl_graph_op *const_dnnl_graph_op_t;
+
+/// @} dnnl_graph_api_op
+
+/// @addtogroup dnnl_graph_api_allocator
+/// @{
+
+/// Allocation call-back function interface for host. For SYCL allocator, see
+/// #dnnl_graph_sycl_allocate_f.
+typedef void *(*dnnl_graph_host_allocate_f)(size_t size, size_t alignment);
+
+/// Deallocation call-back function interface for host. For SYCL allocator, see
+/// #dnnl_graph_sycl_deallocate_f.
+typedef void (*dnnl_graph_host_deallocate_f)(void *);
+
+/// An opaque structure to describe an allocator.
+struct dnnl_graph_allocator;
+
+/// An allocator handle.
+typedef struct dnnl_graph_allocator *dnnl_graph_allocator_t;
+
+/// A constant allocator handle.
+typedef const struct dnnl_graph_allocator *const_dnnl_graph_allocator_t;
+
+/// @} dnnl_graph_api_allocator
+
+/// @addtogroup dnnl_graph_api_compiled_partition
+/// @{
+
+/// In-place pair definition. It can queried from a compiled partition
+/// indicating that an input and an output of the partition can share the same
+/// memory buffer for computation. In-place computation helps to reduce the
+/// memory footprint and improves cache locality. But since the library may not
+/// have a global view of user's application, it's possible that the tensor with
+/// `input_id` is used at other places in user's computation graph. In this
+/// case, the user should take the in-place pair as a hint and pass a different
+/// memory buffer for output tensor to avoid overwriting the input memory buffer
+/// which will probably cause unexpected incorrect results.
+typedef struct {
+    /// The id of input tensor
+    size_t input_id;
+
+    /// The id of output tensor
+    size_t output_id;
+} dnnl_graph_inplace_pair_t;
+
+/// An opaque structure to describe a compiled partition.
+struct dnnl_graph_compiled_partition;
+
+/// A compiled partition handle.
+typedef struct dnnl_graph_compiled_partition *dnnl_graph_compiled_partition_t;
+
+/// A constant compiled partition handle.
+typedef const struct dnnl_graph_compiled_partition
+        *const_dnnl_graph_compiled_partition_t;
+
+/// @} dnnl_graph_api_compiled_partition
+
+/// @addtogroup dnnl_graph_api_tensor
+/// @{
+
+/// An opaque structure to describe a tensor.
+struct dnnl_graph_tensor;
+
+/// A tensor handle.
+typedef struct dnnl_graph_tensor *dnnl_graph_tensor_t;
+
+/// A constant tensor handle.
+typedef const struct dnnl_graph_tensor *const_dnnl_graph_tensor_t;
+
+/// @} dnnl_graph_api_tensor
+
+/// @} dnnl_graph_api
+
+/// @} dnnl_api
+
+#ifdef __cplusplus
+}
+#endif
+#endif

phivenv/Lib/site-packages/torch/include/oneapi/dnnl/dnnl_ocl.h

@@ -0,0 +1,276 @@
+/*******************************************************************************
+* Copyright 2020-2024 Intel Corporation
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*******************************************************************************/
+
+#ifndef ONEAPI_DNNL_DNNL_OCL_H
+#define ONEAPI_DNNL_DNNL_OCL_H
+
+#include "oneapi/dnnl/dnnl.h"
+
+#include "oneapi/dnnl/dnnl_ocl_types.h"
+
+/// @cond DO_NOT_DOCUMENT_THIS
+// Set target version for OpenCL explicitly to suppress a compiler warning.
+#ifndef CL_TARGET_OPENCL_VERSION
+#define CL_TARGET_OPENCL_VERSION 120
+#endif
+
+#include <CL/cl.h>
+/// @endcond
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/// @addtogroup dnnl_api
+/// @{
+
+/// @addtogroup dnnl_api_interop
+/// @{
+
+/// @addtogroup dnnl_api_ocl_interop
+/// @{
+
+/// Creates a memory object.
+///
+/// Unless @p handle is equal to DNNL_MEMORY_NONE or DNNL_MEMORY_ALLOCATE, the
+/// constructed memory object will have the underlying buffer set. In this
+/// case, the buffer will be initialized as if:
+/// - dnnl_memory_set_data_handle() has been called, if @p memory_kind is equal
+///   to dnnl_ocl_interop_usm, or
+/// - dnnl_ocl_interop_memory_set_mem_object() has been called, if @p memory_kind
+///   is equal to dnnl_ocl_interop_buffer.
+///
+/// @param memory Output memory object.
+/// @param memory_desc Memory descriptor.
+/// @param engine Engine to use.
+/// @param memory_kind Memory allocation kind to specify the type of handle.
+/// @param handle Handle of the memory buffer to use as an underlying storage.
+///     - A USM pointer to the user-allocated buffer. In this case the library
+///       doesn't own the buffer. Requires @p memory_kind to be equal to
+///       dnnl_ocl_interop_usm.
+///     - An OpenCL buffer. In this case the library doesn't own the buffer.
+///       Requires @p memory_kind be equal to be equal to dnnl_ocl_interop_buffer.
+///     - The DNNL_MEMORY_ALLOCATE special value. Instructs the library to
+///       allocate the buffer that corresponds to the memory allocation kind
+///       @p memory_kind for the memory object. In this case the library
+///       owns the buffer.
+///     - The DNNL_MEMORY_NONE specific value. Instructs the library to
+///       create memory object without an underlying buffer.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_ocl_interop_memory_create(dnnl_memory_t *memory,
+        const_dnnl_memory_desc_t memory_desc, dnnl_engine_t engine,
+        dnnl_ocl_interop_memory_kind_t memory_kind, void *handle);
+
+#ifdef DNNL_EXPERIMENTAL_SPARSE
+/// Creates a memory object with multiple handles.
+///
+/// @param memory Output memory object.
+/// @param memory_desc Memory descriptor.
+/// @param engine Engine to use.
+/// @param memory_kind Memory allocation kind to specify the type of handles.
+/// @param nhandles Number of handles.
+/// @param handles Handles of the memory buffers to use as underlying storages.
+///     For each element of the @p handles array the following applies:
+///     - A USM pointer to the user-allocated buffer. In this case the library
+///       doesn't own the buffer. Requires @p memory_kind to be equal to
+///       dnnl_ocl_interop_usm.
+///     - An OpenCL buffer. In this case the library doesn't own the buffer.
+///       Requires @p memory_kind be equal to be equal to dnnl_ocl_interop_buffer.
+///     - The DNNL_MEMORY_ALLOCATE special value. Instructs the library to
+///       allocate the buffer that corresponds to the memory allocation kind
+///       @p memory_kind for the memory object. In this case the library
+///       owns the buffer.
+///     - The DNNL_MEMORY_NONE specific value. Instructs the library to
+///       create memory object without an underlying buffer.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_ocl_interop_memory_create_v2(dnnl_memory_t *memory,
+        const_dnnl_memory_desc_t memory_desc, dnnl_engine_t engine,
+        dnnl_ocl_interop_memory_kind_t memory_kind, int nhandles,
+        void **handles);
+#endif
+
+/// Returns the memory allocation kind associated with a memory object.
+///
+/// @param memory Memory to query.
+/// @param memory_kind Output underlying memory allocation kind of the memory
+///     object.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_ocl_interop_memory_get_memory_kind(
+        const_dnnl_memory_t memory,
+        dnnl_ocl_interop_memory_kind_t *memory_kind);
+
+/// Returns an OpenCL memory object associated with a memory object.
+///
+/// @param memory Memory object.
+/// @param mem_object Output OpenCL memory object.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_ocl_interop_memory_get_mem_object(
+        const_dnnl_memory_t memory, cl_mem *mem_object);
+
+/// Sets OpenCL memory object associated with a memory object.
+///
+/// For behavioral details, see dnnl_memory_set_data_handle().
+///
+/// @param memory Memory object.
+/// @param mem_object OpenCL memory object.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_ocl_interop_memory_set_mem_object(
+        dnnl_memory_t memory, cl_mem mem_object);
+
+/// Retrieves a cache blob ID for the OpenCL device.
+///
+/// @warning
+///     This API is intended to be used with
+///     #dnnl_ocl_interop_engine_get_cache_blob() and
+///     #dnnl_ocl_interop_engine_create_from_cache_blob(). The returned cache
+///     blob ID can only be used as an ID of the cache blob returned by
+///     #dnnl_ocl_interop_engine_get_cache_blob().
+///
+/// @note The cache blob ID can be empty (@p size will be 0 and
+///     @p cache_blob_id will be nullptr) if oneDNN doesn't have anything to
+///     put in the cache blob. (#dnnl_ocl_interop_engine_get_cache_blob will
+///     return an empty cache blob).
+///
+/// @param device An OpenCL device.
+/// @param size Size of the cache blob ID in bytes.
+/// @param cache_blob_id Cache blob id of size @p size. If
+///     the @p cache_blob_id is nullptr then the size of the cache blob ID is
+///     returned in @p size.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_ocl_interop_engine_get_cache_blob_id(
+        cl_device_id device, size_t *size, uint8_t *cache_blob_id);
+
+/// Retrieves a cache blob associated with the given engine.
+///
+/// @note The cache blob can be empty (@p size will be 0 and @p cache_blob
+///     will be nullptr) if oneDNN doesn't have anything to put in the cache
+///     blob. It's the user's responsibility to check whether it's empty
+///     prior to passing it to
+///     #dnnl_ocl_interop_engine_create_from_cache_blob().
+///
+/// @param engine Engine to query for the cache blob.
+/// @param size Size of the cache blob in bytes.
+/// @param cache_blob Cache blob of size @p size. If the @p cache_blob is
+///     nullptr then the size of the cache blob is returned in @p size.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_ocl_interop_engine_get_cache_blob(
+        dnnl_engine_t engine, size_t *size, uint8_t *cache_blob);
+
+/// Creates an engine from the given cache blob.
+///
+/// @param engine Output engine.
+/// @param device The OpenCL device that this engine will encapsulate.
+/// @param context The OpenCL context (containing the device) that this
+///     engine will use for all operations.
+/// @param size Size of the cache blob in bytes.
+/// @param cache_blob Cache blob of size @p size.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_ocl_interop_engine_create_from_cache_blob(
+        dnnl_engine_t *engine, cl_device_id device, cl_context context,
+        size_t size, const uint8_t *cache_blob);
+
+/// Creates an engine associated with an OpenCL device and an OpenCL context.
+///
+/// @param engine Output engine.
+/// @param device Underlying OpenCL device to use for the engine.
+/// @param context Underlying OpenCL context to use for the engine.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_ocl_interop_engine_create(
+        dnnl_engine_t *engine, cl_device_id device, cl_context context);
+
+/// Returns the OpenCL context associated with an engine.
+///
+/// @param engine Engine to query.
+/// @param context Output underlying OpenCL context of the engine.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_ocl_interop_engine_get_context(
+        dnnl_engine_t engine, cl_context *context);
+
+/// Returns the OpenCL device associated with an engine.
+///
+/// @param engine Engine to query.
+/// @param device Output underlying OpenCL device of the engine.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_ocl_interop_get_device(
+        dnnl_engine_t engine, cl_device_id *device);
+
+/// Creates an execution stream for a given engine associated with
+/// an OpenCL command queue.
+///
+/// @param stream Output execution stream.
+/// @param engine Engine to create the execution stream on.
+/// @param queue OpenCL command queue to use.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_ocl_interop_stream_create(
+        dnnl_stream_t *stream, dnnl_engine_t engine, cl_command_queue queue);
+
+/// Returns the OpenCL command queue associated with an execution stream.
+///
+/// @param stream Execution stream to query.
+/// @param queue Output OpenCL command queue.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_ocl_interop_stream_get_command_queue(
+        dnnl_stream_t stream, cl_command_queue *queue);
+
+/// Executes computations specified by the primitive in a specified stream and
+/// returns an OpenCL event.
+///
+/// @param primitive Primitive to execute.
+/// @param stream Stream to use.
+/// @param nargs Number of arguments.
+/// @param args Array of arguments. Each argument is an
+///     <index, #dnnl_memory_t> pair. The index is one of the `DNNL_ARG_*`
+///     values such as `DNNL_ARG_SRC`. Unless runtime shapes are used (see
+///     #DNNL_RUNTIME_DIM_VAL), the memory object must have the same memory
+///     descriptor as that returned by
+///     #dnnl_primitive_desc_query_md(#dnnl_query_exec_arg_md, index).
+/// @param deps A pointer to a vector of size @p ndeps that contains
+///     dependencies.
+/// @param ndeps Number of dependencies.
+/// @param return_event Output event. It's the user's responsibility to
+///     manage lifetime of the event. Can be NULL. When @p stream is in-order
+///     NULL will be returned.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_ocl_interop_primitive_execute(
+        const_dnnl_primitive_t primitive, dnnl_stream_t stream, int nargs,
+        const dnnl_exec_arg_t *args, const cl_event *deps, int ndeps,
+        cl_event *return_event);
+
+/// @} dnnl_api_ocl_interop
+
+/// @} dnnl_api_interop
+
+/// @} dnnl_api
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif

phivenv/Lib/site-packages/torch/include/oneapi/dnnl/dnnl_ocl.hpp

@@ -0,0 +1,445 @@
+/*******************************************************************************
+* Copyright 2020-2025 Intel Corporation
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*******************************************************************************/
+
+#ifndef ONEAPI_DNNL_DNNL_OCL_HPP
+#define ONEAPI_DNNL_DNNL_OCL_HPP
+
+#include "oneapi/dnnl/dnnl.hpp"
+
+/// @cond DO_NOT_DOCUMENT_THIS
+#include <algorithm>
+#include <cstdlib>
+#include <iterator>
+#include <memory>
+#include <string>
+#include <vector>
+#include <unordered_map>
+
+#include "oneapi/dnnl/dnnl_ocl.h"
+
+#include <CL/cl.h>
+/// @endcond
+
+/// @addtogroup dnnl_api
+/// @{
+
+namespace dnnl {
+
+/// @addtogroup dnnl_api_interop Runtime interoperability API
+/// API extensions to interact with the underlying run-time.
+/// @{
+
+/// @addtogroup dnnl_api_ocl_interop OpenCL interoperability API
+/// API extensions to interact with the underlying OpenCL run-time.
+///
+/// @sa @ref dev_guide_opencl_interoperability in developer guide
+/// @{
+
+/// OpenCL interoperability namespace
+namespace ocl_interop {
+
+/// Memory allocation kind.
+enum class memory_kind {
+    /// USM (device, shared, host, or unknown) memory allocation kind.
+    usm = dnnl_ocl_interop_usm,
+    /// Buffer memory allocation kind - default.
+    buffer = dnnl_ocl_interop_buffer,
+};
+
+/// Converts a memory allocation kind enum value from C++ API to C API type.
+///
+/// @param akind C++ API memory allocation kind enum value.
+/// @returns Corresponding C API memory allocation kind enum value.
+inline dnnl_ocl_interop_memory_kind_t convert_to_c(memory_kind akind) {
+    return static_cast<dnnl_ocl_interop_memory_kind_t>(akind);
+}
+
+/// Returns the cache blob ID of the OpenCL device.
+///
+/// @warning
+///     This API is intended to be used with
+///     #dnnl::ocl_interop::get_engine_cache_blob() and
+///     #dnnl::ocl_interop::make_engine(cl_device_id, cl_context, const std::vector<uint8_t> &).
+///     The returned cache blob ID can only be used as an ID of the cache blob
+///     returned by #dnnl::ocl_interop::get_engine_cache_blob().
+///
+/// @note The cache blob ID can be empty (@p size will be 0 and
+///     @p cache_blob_id will be nullptr) if oneDNN doesn't have anything to
+///     put in the cache blob. (#dnnl_ocl_interop_engine_get_cache_blob will
+///     return an empty cache blob).
+///
+/// @param device An OpenCL device.
+/// @returns A vector containing the cache blob ID.
+inline std::vector<uint8_t> get_engine_cache_blob_id(cl_device_id device) {
+    size_t size = 0;
+    error::wrap_c_api(
+            dnnl_ocl_interop_engine_get_cache_blob_id(device, &size, nullptr),
+            "could not get an engine cache blob id size");
+
+    std::vector<uint8_t> cache_blob_id(size);
+    error::wrap_c_api(dnnl_ocl_interop_engine_get_cache_blob_id(
+                              device, &size, cache_blob_id.data()),
+            "could not get an engine cache blob id");
+    return cache_blob_id;
+}
+
+/// Returns a cache blob for the engine.
+///
+/// @note The cache blob vector can be empty if oneDNN doesn't have anything
+///     to put in the cache blob. It's the user's responsibility to check
+///     whether it's empty prior to passing it to
+///     #dnnl::ocl_interop::make_engine(cl_device_id, cl_context, const std::vector<uint8_t> &)
+///
+/// @param aengine Engine to query for the cache blob.
+/// @returns Vector containing the cache blob.
+inline std::vector<uint8_t> get_engine_cache_blob(const engine &aengine) {
+    size_t size = 0;
+    error::wrap_c_api(dnnl_ocl_interop_engine_get_cache_blob(
+                              aengine.get(), &size, nullptr),
+            "could not get an engine cache blob size");
+
+    std::vector<uint8_t> cache_blob(size);
+    error::wrap_c_api(dnnl_ocl_interop_engine_get_cache_blob(
+                              aengine.get(), &size, cache_blob.data()),
+            "could not get an engine cache blob");
+    return cache_blob;
+}
+
+/// Constructs an engine from the given cache blob.
+///
+/// @param device The OpenCL device that this engine will encapsulate.
+/// @param context The OpenCL context (containing the device) that this
+///     engine will use for all operations.
+/// @param cache_blob Cache blob.
+/// @returns An engine.
+inline engine make_engine(cl_device_id device, cl_context context,
+        const std::vector<uint8_t> &cache_blob) {
+    dnnl_engine_t c_engine;
+    error::wrap_c_api(
+            dnnl_ocl_interop_engine_create_from_cache_blob(&c_engine, device,
+                    context, cache_blob.size(), cache_blob.data()),
+            "could not create an engine from cache blob");
+    return engine(c_engine);
+}
+
+/// Constructs an engine from OpenCL device and context objects.
+///
+/// @param device The OpenCL device that this engine will encapsulate.
+/// @param context The OpenCL context (containing the device) that this
+///     engine will use for all operations.
+/// @returns An engine.
+inline engine make_engine(cl_device_id device, cl_context context) {
+    dnnl_engine_t c_engine;
+    error::wrap_c_api(
+            dnnl_ocl_interop_engine_create(&c_engine, device, context),
+            "could not create an engine");
+    return engine(c_engine);
+}
+
+/// Returns OpenCL context associated with the engine.
+///
+/// @param aengine An engine.
+/// @returns Underlying OpenCL context.
+inline cl_context get_context(const engine &aengine) {
+    cl_context context = nullptr;
+    error::wrap_c_api(
+            dnnl_ocl_interop_engine_get_context(aengine.get(), &context),
+            "could not get an OpenCL context from an engine");
+    return context;
+}
+
+/// Returns OpenCL device associated with the engine.
+///
+/// @param aengine An engine.
+/// @returns Underlying OpenCL device.
+inline cl_device_id get_device(const engine &aengine) {
+    cl_device_id device = nullptr;
+    error::wrap_c_api(dnnl_ocl_interop_get_device(aengine.get(), &device),
+            "could not get an OpenCL device from an engine");
+    return device;
+}
+
+/// Constructs an execution stream for the specified engine and OpenCL queue.
+///
+/// @param aengine Engine to create the stream on.
+/// @param queue OpenCL queue to use for the stream.
+/// @returns An execution stream.
+inline stream make_stream(const engine &aengine, cl_command_queue queue) {
+    dnnl_stream_t c_stream;
+    error::wrap_c_api(
+            dnnl_ocl_interop_stream_create(&c_stream, aengine.get(), queue),
+            "could not create a stream");
+    return stream(c_stream);
+}
+
+/// Returns OpenCL queue object associated with the execution stream.
+///
+/// @param astream An execution stream.
+/// @returns Underlying OpenCL queue.
+inline cl_command_queue get_command_queue(const stream &astream) {
+    cl_command_queue queue = nullptr;
+    error::wrap_c_api(
+            dnnl_ocl_interop_stream_get_command_queue(astream.get(), &queue),
+            "could not get an OpenCL command queue from a stream");
+    return queue;
+}
+
+/// Returns the OpenCL memory object associated with the memory object.
+///
+/// @param amemory A memory object.
+/// @returns Underlying OpenCL memory object.
+inline cl_mem get_mem_object(const memory &amemory) {
+    cl_mem mem_object;
+    error::wrap_c_api(
+            dnnl_ocl_interop_memory_get_mem_object(amemory.get(), &mem_object),
+            "could not get OpenCL buffer object from a memory object");
+    return mem_object;
+}
+
+/// Sets the OpenCL memory object associated with the memory object.
+///
+/// For behavioral details see memory::set_data_handle().
+///
+/// @param amemory A memory object.
+/// @param mem_object OpenCL cl_mem object to use as the underlying
+///     storage. It must have at least get_desc().get_size() bytes
+///     allocated.
+inline void set_mem_object(memory &amemory, cl_mem mem_object) {
+    error::wrap_c_api(
+            dnnl_ocl_interop_memory_set_mem_object(amemory.get(), mem_object),
+            "could not set OpenCL buffer object from a memory object");
+}
+
+/// Returns the memory allocation kind associated with a memory object.
+///
+/// @param amemory A memory object.
+///
+/// @returns The underlying memory allocation kind of the memory object.
+inline memory_kind get_memory_kind(const memory &amemory) {
+    dnnl_ocl_interop_memory_kind_t ckind;
+    error::wrap_c_api(
+            dnnl_ocl_interop_memory_get_memory_kind(amemory.get(), &ckind),
+            "could not get memory kind");
+    return static_cast<memory_kind>(ckind);
+}
+
+#ifdef DNNL_EXPERIMENTAL_SPARSE
+/// Creates a memory object with multiple handles.
+///
+/// @param memory_desc Memory descriptor.
+/// @param aengine Engine to use.
+/// @param kind Memory allocation kind to specify the type of handles.
+/// @param handles Handles of the memory buffers to use as underlying storages.
+///     For each element of the @p handles array the following applies:
+///     - A USM pointer to the user-allocated buffer. In this case the library
+///       doesn't own the buffer. Requires @p memory_kind to be equal to
+///       dnnl_ocl_interop_usm.
+///     - An OpenCL buffer. In this case the library doesn't own the buffer.
+///       Requires @p memory_kind be equal to be equal to dnnl_ocl_interop_buffer.
+///     - The DNNL_MEMORY_ALLOCATE special value. Instructs the library to
+///       allocate the buffer that corresponds to the memory allocation kind
+///       @p memory_kind for the memory object. In this case the library
+///       owns the buffer.
+///     - The DNNL_MEMORY_NONE specific value. Instructs the library to
+///       create memory object without an underlying buffer.
+///
+///  If the @p handles vector is not provided the library will allocate all
+///  buffers as if all handles have the special value DNNL_MEMORY_ALLOCATE.
+///
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+inline memory make_memory(const memory::desc &memory_desc,
+        const engine &aengine, memory_kind kind,
+        std::vector<void *> handles = {}) {
+    if (handles.empty()) {
+        const int nhandles = memory_desc.get_num_handles();
+        handles.resize(nhandles, DNNL_MEMORY_ALLOCATE);
+    }
+
+    dnnl_memory_t c_memory;
+    error::wrap_c_api(
+            dnnl_ocl_interop_memory_create_v2(&c_memory, memory_desc.get(),
+                    aengine.get(), convert_to_c(kind), (int)handles.size(),
+                    handles.data()),
+            "could not create a memory");
+    return memory(c_memory);
+}
+
+/// Constructs a memory object with multiple OpenCL buffers.
+///
+/// @param memory_desc Memory descriptor.
+/// @param aengine Engine to use.
+/// @param mem_objects A vector of OpenCL buffers to use.
+///
+/// @returns Created memory object.
+inline memory make_memory(const memory::desc &memory_desc,
+        const engine &aengine, std::vector<cl_mem> mem_objects) {
+    const int nhandles = memory_desc.get_num_handles();
+    std::vector<void *> handles(nhandles, DNNL_MEMORY_NONE);
+    memory amemory(memory_desc, aengine, handles);
+    for (int i = 0; i < nhandles; i++)
+        amemory.set_data_handle(mem_objects[i], i);
+    return amemory;
+}
+
+/// Creates a memory object.
+///
+/// Unless @p handle is equal to DNNL_MEMORY_NONE or DNNL_MEMORY_ALLOCATE, the
+/// constructed memory object will have the underlying buffer set. In this
+/// case, the buffer will be initialized as if:
+/// - dnnl::memory::set_data_handle() had been called, if @p memory_kind is
+///   equal to dnnl::ocl_interop::memory_kind::usm, or
+/// - dnnl::ocl_interop::set_mem_object() has been called, if @p memory_kind is
+///   equal to dnnl::ocl_interop::memory_kind::buffer.
+///
+/// @param memory_desc Memory descriptor.
+/// @param aengine Engine to use.
+/// @param kind Memory allocation kind to specify the type of handle.
+/// @param handle Handle of the memory buffer to use as an underlying storage.
+///     - A USM pointer to the user-allocated buffer. In this case the library
+///       doesn't own the buffer. Requires @p memory_kind to be equal to
+///       dnnl::ocl_interop::memory_kind::usm.
+///     - An OpenCL buffer. In this case the library doesn't own the buffer.
+///       Requires @p memory_kind be equal to be equal to
+///       dnnl::ocl_interop::memory_kind::buffer.
+///     - The DNNL_MEMORY_ALLOCATE special value. Instructs the library to
+///       allocate the buffer that corresponds to the memory allocation kind
+///       @p memory_kind for the memory object. In this case the library
+///       owns the buffer.
+///     - The DNNL_MEMORY_NONE specific value. Instructs the library to
+///       create memory object without an underlying buffer.
+///
+/// @returns Created memory object.
+inline memory make_memory(const memory::desc &memory_desc,
+        const engine &aengine, memory_kind kind, void *handle) {
+    return make_memory(
+            memory_desc, aengine, kind, std::vector<void *> {handle});
+}
+
+/// Constructs a memory object from an OpenCL buffer.
+///
+/// @param memory_desc Memory descriptor.
+/// @param aengine Engine to use.
+/// @param mem_object An OpenCL buffer to use.
+///
+/// @returns Created memory object.
+inline memory make_memory(const memory::desc &memory_desc,
+        const engine &aengine, cl_mem mem_object) {
+    return make_memory(memory_desc, aengine, std::vector<cl_mem> {mem_object});
+}
+#else
+
+/// Creates a memory object.
+///
+/// Unless @p handle is equal to DNNL_MEMORY_NONE or DNNL_MEMORY_ALLOCATE, the
+/// constructed memory object will have the underlying buffer set. In this
+/// case, the buffer will be initialized as if:
+/// - dnnl::memory::set_data_handle() had been called, if @p memory_kind is
+///   equal to dnnl::ocl_interop::memory_kind::usm, or
+/// - dnnl::ocl_interop::set_mem_object() has been called, if @p memory_kind is
+///   equal to dnnl::ocl_interop::memory_kind::buffer.
+///
+/// @param memory_desc Memory descriptor.
+/// @param aengine Engine to use.
+/// @param kind Memory allocation kind to specify the type of handle.
+/// @param handle Handle of the memory buffer to use as an underlying storage.
+///     - A USM pointer to the user-allocated buffer. In this case the library
+///       doesn't own the buffer. Requires @p memory_kind to be equal to
+///       dnnl::ocl_interop::memory_kind::usm.
+///     - An OpenCL buffer. In this case the library doesn't own the buffer.
+///       Requires @p memory_kind be equal to be equal to
+///       dnnl::ocl_interop::memory_kind::buffer.
+///     - The DNNL_MEMORY_ALLOCATE special value. Instructs the library to
+///       allocate the buffer that corresponds to the memory allocation kind
+///       @p memory_kind for the memory object. In this case the library
+///       owns the buffer.
+///     - The DNNL_MEMORY_NONE specific value. Instructs the library to
+///       create memory object without an underlying buffer.
+///
+/// @returns Created memory object.
+inline memory make_memory(const memory::desc &memory_desc,
+        const engine &aengine, memory_kind kind,
+        void *handle = DNNL_MEMORY_ALLOCATE) {
+    dnnl_memory_t c_memory;
+    error::wrap_c_api(
+            dnnl_ocl_interop_memory_create(&c_memory, memory_desc.get(),
+                    aengine.get(), convert_to_c(kind), handle),
+            "could not create a memory");
+    return memory(c_memory);
+}
+
+/// Constructs a memory object from an OpenCL buffer.
+///
+/// @param memory_desc Memory descriptor.
+/// @param aengine Engine to use.
+/// @param mem_object An OpenCL buffer to use.
+///
+/// @returns Created memory object.
+inline memory make_memory(const memory::desc &memory_desc,
+        const engine &aengine, cl_mem mem_object) {
+    memory amemory(memory_desc, aengine, DNNL_MEMORY_NONE);
+    set_mem_object(amemory, mem_object);
+    return amemory;
+}
+#endif
+
+/// Executes computations specified by the primitive in a specified stream and
+/// returns a SYCL event.
+///
+/// Arguments are passed via an arguments map containing
+/// <index, memory object> pairs. The index must be one of the `DNNL_ARG_*`
+/// values such as `DNNL_ARG_SRC`, and the memory must have a memory descriptor
+/// matching the one returned by
+/// #dnnl::primitive_desc::query_md(#query::exec_arg_md, index) unless using
+/// dynamic shapes (see #DNNL_RUNTIME_DIM_VAL).
+///
+/// @param aprimitive Primitive to execute.
+/// @param astream Stream object. The stream must belong to the same engine
+///     as the primitive.
+/// @param args Arguments map.
+/// @param deps Optional vector with `cl_event` dependencies.
+///
+/// @returns Output event. It's the user's responsibility to manage lifetime
+///     of the event.
+inline cl_event execute(const dnnl::primitive &aprimitive,
+        const stream &astream, const std::unordered_map<int, memory> &args,
+        const std::vector<cl_event> &deps = {}) {
+    std::vector<dnnl_exec_arg_t> c_args;
+    c_args.reserve(args.size());
+    for (const auto &a : args)
+        c_args.push_back({a.first, a.second.get()});
+
+    const cl_event *c_deps = deps.empty() ? nullptr : deps.data();
+
+    cl_event return_event;
+    error::wrap_c_api(dnnl_ocl_interop_primitive_execute(aprimitive.get(),
+                              astream.get(), (int)c_args.size(), c_args.data(),
+                              c_deps, (int)deps.size(), &return_event),
+            "could not execute a primitive");
+    return return_event;
+}
+
+} // namespace ocl_interop
+
+/// @} dnnl_api_ocl_interop
+
+/// @} dnnl_api_interop
+
+} // namespace dnnl
+
+/// @} dnnl_api
+
+#endif

phivenv/Lib/site-packages/torch/include/oneapi/dnnl/dnnl_ocl_types.h

@@ -0,0 +1,51 @@
+/*******************************************************************************
+* Copyright 2021 Intel Corporation
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*******************************************************************************/
+
+#ifndef ONEAPI_DNNL_DNNL_OCL_TYPES_H
+#define ONEAPI_DNNL_DNNL_OCL_TYPES_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/// @addtogroup dnnl_api
+/// @{
+
+/// @addtogroup dnnl_api_interop
+/// @{
+
+/// @addtogroup dnnl_api_ocl_interop
+/// @{
+
+/// Memory allocation kind.
+typedef enum {
+    /// USM (device, shared, host, or unknown) memory allocation kind.
+    dnnl_ocl_interop_usm,
+    /// Buffer memory allocation kind - default.
+    dnnl_ocl_interop_buffer,
+} dnnl_ocl_interop_memory_kind_t;
+
+/// @} dnnl_api_ocl_interop
+
+/// @} dnnl_api_interop
+
+/// @} dnnl_api
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif

phivenv/Lib/site-packages/torch/include/oneapi/dnnl/dnnl_sycl.h

@@ -0,0 +1,199 @@
+/*******************************************************************************
+* Copyright 2020-2024 Intel Corporation
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*******************************************************************************/
+
+#ifndef ONEAPI_DNNL_DNNL_SYCL_H
+#define ONEAPI_DNNL_DNNL_SYCL_H
+
+#include "oneapi/dnnl/dnnl.h"
+
+#include "oneapi/dnnl/dnnl_sycl_types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/// @addtogroup dnnl_api
+/// @{
+
+/// @addtogroup dnnl_api_interop
+/// @{
+
+/// @addtogroup dnnl_api_sycl_interop
+/// @{
+
+/// Creates an engine associated with a SYCL device and a SYCL context.
+///
+/// @param engine Output engine.
+/// @param device Pointer to the SYCL device to use for the engine.
+/// @param context Pointer to the SYCL context to use for the engine.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_sycl_interop_engine_create(
+        dnnl_engine_t *engine, const void *device, const void *context);
+
+/// Returns the SYCL context associated with an engine.
+///
+/// @param engine Engine to query.
+/// @param context Pointer to the underlying SYCL context of the engine.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_sycl_interop_engine_get_context(
+        dnnl_engine_t engine, void **context);
+
+/// Returns the SYCL device associated with an engine.
+///
+/// @param engine Engine to query.
+/// @param device Pointer to the underlying SYCL device of the engine.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_sycl_interop_engine_get_device(
+        dnnl_engine_t engine, void **device);
+
+/// Creates a memory object.
+///
+/// Unless @p handle is equal to DNNL_MEMORY_NONE or DNNL_MEMORY_ALLOCATE, the
+/// constructed memory object will have the underlying buffer set. In this
+/// case, the buffer will be initialized as if:
+/// - dnnl_memory_set_data_handle() had been called, if @p memory_kind is equal
+///   to dnnl_sycl_interop_usm, or
+/// - dnnl_sycl_interop_memory_set_buffer() has been called, if @p memory_kind
+///   is equal to dnnl_sycl_interop_buffer.
+///
+/// @param memory Output memory object.
+/// @param memory_desc Memory descriptor.
+/// @param engine Engine to use.
+/// @param memory_kind Memory allocation kind to specify the type of handle.
+/// @param handle Handle of the memory buffer to use as an underlying storage.
+///     - A USM pointer to the user-allocated buffer. In this case the library
+///       doesn't own the buffer. Requires @p memory_kind to be equal to
+///       dnnl_sycl_interop_usm.
+///     - A pointer to SYCL buffer. In this case the library doesn't own the
+///       buffer. Requires @p memory_kind be equal to be equal to
+///       dnnl_sycl_interop_buffer.
+///     - The DNNL_MEMORY_ALLOCATE special value. Instructs the library to
+///       allocate the buffer that corresponds to the memory allocation kind
+///       @p memory_kind for the memory object. In this case the library
+///       owns the buffer.
+///     - The DNNL_MEMORY_NONE specific value. Instructs the library to
+///       create memory object without an underlying buffer.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_sycl_interop_memory_create(dnnl_memory_t *memory,
+        const_dnnl_memory_desc_t memory_desc, dnnl_engine_t engine,
+        dnnl_sycl_interop_memory_kind_t memory_kind, void *handle);
+
+#ifdef DNNL_EXPERIMENTAL_SPARSE
+/// Creates a memory object with multiple handles.
+///
+/// @param memory Output memory object.
+/// @param memory_desc Memory descriptor.
+/// @param engine Engine to use.
+/// @param memory_kind Memory allocation kind to specify the type of handles.
+/// @param nhandles Number of handles.
+/// @param handles Handles of the memory buffers to use as underlying storages.
+///     For each element of the @p handles array the following applies:
+///     - A USM pointer to the user-allocated buffer. In this case the library
+///       doesn't own the buffer. Requires @p memory_kind to be equal to
+///       dnnl_sycl_interop_usm.
+///     - A pointer to SYCL buffer. In this case the library doesn't own the
+///       buffer. Requires @p memory_kind be equal to be equal to
+///       dnnl_sycl_interop_buffer.
+///     - The DNNL_MEMORY_ALLOCATE special value. Instructs the library to
+///       allocate the buffer that corresponds to the memory allocation kind
+///       @p memory_kind for the memory object. In this case the library
+///       owns the buffer.
+///     - The DNNL_MEMORY_NONE specific value. Instructs the library to
+///       create memory object without an underlying buffer.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_sycl_interop_memory_create_v2(dnnl_memory_t *memory,
+        const_dnnl_memory_desc_t memory_desc, dnnl_engine_t engine,
+        dnnl_sycl_interop_memory_kind_t memory_kind, int nhandles,
+        void **handles);
+#endif
+
+/// Returns the memory allocation kind associated with a memory object.
+///
+/// @param memory Memory to query.
+/// @param memory_kind Output underlying memory allocation kind of the memory
+///     object.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_sycl_interop_memory_get_memory_kind(
+        const_dnnl_memory_t memory,
+        dnnl_sycl_interop_memory_kind_t *memory_kind);
+
+/// Sets a SYCL buffer for a memory object.
+///
+/// @param memory Memory object.
+/// @param buffer SYCL buffer to be set in the memory object.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_sycl_interop_memory_set_buffer(
+        dnnl_memory_t memory, void *buffer);
+
+/// Creates an execution stream for a given engine associated with a SYCL
+/// queue.
+///
+/// @param stream Output execution stream.
+/// @param engine Engine to create the execution stream on.
+/// @param queue SYCL queue to use.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_sycl_interop_stream_create(
+        dnnl_stream_t *stream, dnnl_engine_t engine, void *queue);
+
+/// Returns the SYCL queue associated with an execution stream.
+///
+/// @param stream Execution stream to query.
+/// @param queue Output SYCL command queue.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_sycl_interop_stream_get_queue(
+        dnnl_stream_t stream, void **queue);
+
+/// Executes computations specified by the primitive in a specified stream and
+/// returns a SYCL event.
+///
+/// @param primitive Primitive to execute.
+/// @param stream Stream to use.
+/// @param nargs Number of arguments.
+/// @param args Array of arguments. Each argument is an
+///     <index, #dnnl_memory_t> pair. The index is one of the `DNNL_ARG_*`
+///     values such as `DNNL_ARG_SRC`. Unless runtime shapes are used (see
+///     #DNNL_RUNTIME_DIM_VAL), the memory object must have the same memory
+///     descriptor as that returned by
+///     #dnnl_primitive_desc_query_md(#dnnl_query_exec_arg_md, index).
+/// @param deps A pointer to std::vector<sycl::event> that contains
+///     dependencies.
+/// @param return_event Output event.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_sycl_interop_primitive_execute(
+        const_dnnl_primitive_t primitive, dnnl_stream_t stream, int nargs,
+        const dnnl_exec_arg_t *args, const void *deps, void *return_event);
+
+/// @} dnnl_api_sycl_interop
+
+/// @} dnnl_api_interop
+
+/// @} dnnl_api
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif

phivenv/Lib/site-packages/torch/include/oneapi/dnnl/dnnl_sycl.hpp

@@ -0,0 +1,384 @@
+/*******************************************************************************
+* Copyright 2020-2025 Intel Corporation
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*******************************************************************************/
+
+#ifndef ONEAPI_DNNL_DNNL_SYCL_HPP
+#define ONEAPI_DNNL_DNNL_SYCL_HPP
+
+/// @cond DO_NOT_DOCUMENT_THIS
+#include <algorithm>
+#include <cstdlib>
+#include <iterator>
+#include <memory>
+#include <string>
+#include <vector>
+#include <unordered_map>
+
+#if __has_include(<sycl/sycl.hpp>)
+#include <sycl/sycl.hpp>
+#else
+#error "Unsupported compiler"
+#endif
+
+#include "oneapi/dnnl/dnnl.hpp"
+#include "oneapi/dnnl/dnnl_sycl.h"
+
+/// @endcond
+
+/// @addtogroup dnnl_api
+/// @{
+
+namespace dnnl {
+
+/// @addtogroup dnnl_api_interop
+/// @{
+
+/// @addtogroup dnnl_api_sycl_interop SYCL interoperability API
+/// API extensions to interact with the underlying SYCL run-time.
+///
+/// @sa @ref dev_guide_dpcpp_interoperability in developer guide
+/// @{
+
+/// SYCL interoperability namespace
+namespace sycl_interop {
+
+/// Memory allocation kind.
+enum class memory_kind {
+    /// USM (device, shared, host, or unknown) memory allocation kind - default.
+    usm = dnnl_sycl_interop_usm,
+    /// Buffer memory allocation kind.
+    buffer = dnnl_sycl_interop_buffer,
+};
+
+/// Converts a memory allocation kind enum value from C++ API to C API type.
+///
+/// @param akind C++ API memory allocation kind enum value.
+/// @returns Corresponding C API memory allocation kind enum value.
+inline dnnl_sycl_interop_memory_kind_t convert_to_c(memory_kind akind) {
+    return static_cast<dnnl_sycl_interop_memory_kind_t>(akind);
+}
+
+/// Constructs an engine from SYCL device and context objects.
+///
+/// @param adevice SYCL device.
+/// @param acontext SYCL context.
+///
+/// @returns Created engine.
+inline engine make_engine(
+        const sycl::device &adevice, const sycl::context &acontext) {
+    dnnl_engine_t aengine;
+    error::wrap_c_api(dnnl_sycl_interop_engine_create(&aengine,
+                              static_cast<const void *>(&adevice),
+                              static_cast<const void *>(&acontext)),
+            "could not create an engine");
+    return engine(aengine);
+}
+
+/// Returns the SYCL context associated with an engine.
+///
+/// @param aengine Engine to query.
+///
+/// @returns The underlying SYCL device of the engine.
+inline sycl::context get_context(const engine &aengine) {
+    void *ctx_ptr;
+    error::wrap_c_api(
+            dnnl_sycl_interop_engine_get_context(aengine.get(), &ctx_ptr),
+            "could not get a context handle");
+    auto ctx = *static_cast<sycl::context *>(ctx_ptr);
+    return ctx;
+}
+
+/// Returns the SYCL device associated with an engine.
+///
+/// @param aengine Engine to query.
+///
+/// @returns The underlying SYCL context of the engine.
+inline sycl::device get_device(const engine &aengine) {
+    void *dev_ptr;
+    error::wrap_c_api(
+            dnnl_sycl_interop_engine_get_device(aengine.get(), &dev_ptr),
+            "could not get a device handle");
+    auto dev = *static_cast<sycl::device *>(dev_ptr);
+    return dev;
+}
+
+/// Creates an execution stream for a given engine associated with a SYCL
+/// queue.
+///
+/// @param aengine Engine object to use for the stream.
+/// @param aqueue SYCL queue to use for the stream.
+///
+/// @returns An execution stream.
+inline stream make_stream(const engine &aengine, sycl::queue &aqueue) {
+    dnnl_stream_t astream;
+    error::wrap_c_api(
+            dnnl_sycl_interop_stream_create(&astream, aengine.get(), &aqueue),
+            "could not create a stream");
+    return stream(astream);
+}
+
+/// Returns the SYCL queue associated with an execution stream.
+///
+/// @param astream Execution stream to query.
+///
+/// @returns SYCL queue object.
+inline sycl::queue get_queue(const stream &astream) {
+    void *queue_ptr;
+    error::wrap_c_api(
+            dnnl_sycl_interop_stream_get_queue(astream.get(), &queue_ptr),
+            "could not get a stream handle");
+    auto queue = *static_cast<sycl::queue *>(queue_ptr);
+    return queue;
+}
+
+/// Returns the SYCL buffer associated with a memory object.
+///
+/// Throws an exception if the memory allocation kind associated with the
+/// memory object is not equal to dnnl::sycl_interop::memory_kind::buffer.
+///
+/// @tparam T Type of the requested buffer.
+/// @tparam ndims Number of dimensions of the requested buffer.
+/// @param amemory Memory object.
+///
+/// @returns SYCL buffer associated with the memory object.
+template <typename T, int ndims = 1>
+sycl::buffer<T, ndims> get_buffer(const memory &amemory) {
+    static_assert(ndims == 1, "only 1D buffers supported");
+
+    // XXX: workaround: when CPU runtime is not SYCL and amemory was created
+    // for CPU engine `get_buffer` should return an error. Use interop API to
+    // implement the check.
+    dnnl_sycl_interop_memory_kind_t ckind;
+    error::wrap_c_api(
+            dnnl_sycl_interop_memory_get_memory_kind(amemory.get(), &ckind),
+            "could not get SYCL buffer object");
+
+    void *handle_ptr;
+    error::wrap_c_api(dnnl_memory_get_data_handle(amemory.get(), &handle_ptr),
+            "could not get SYCL buffer object");
+
+    // XXX: workaround: zero-range buffer cannot be constructed.
+    if (!handle_ptr) return sycl::buffer<T, ndims>(sycl::range<1>(1));
+
+    auto &buf_u8 = *static_cast<sycl::buffer<uint8_t, 1> *>(handle_ptr);
+
+    auto range = sycl::range<1>(buf_u8.byte_size() / sizeof(T));
+    return buf_u8.reinterpret<T, 1>(range);
+}
+
+/// Sets SYCL buffer associated with a memory object.
+///
+/// @tparam T Type of the buffer.
+/// @tparam ndims Number of dimensions of the buffer.
+/// @param amemory Memory object to change.
+/// @param abuffer SYCL buffer.
+template <typename T, int ndims>
+void set_buffer(memory &amemory, sycl::buffer<T, ndims> &abuffer) {
+    auto range = sycl::range<1>(abuffer.byte_size());
+    auto buf_u8 = abuffer.template reinterpret<uint8_t, 1>(range);
+    error::wrap_c_api(dnnl_sycl_interop_memory_set_buffer(
+                              amemory.get(), static_cast<void *>(&buf_u8)),
+            "could not set SYCL buffer object");
+}
+
+/// Returns the memory allocation kind associated with a memory object.
+///
+/// @param amemory A memory object.
+///
+/// @returns The underlying memory allocation kind of the memory object.
+inline memory_kind get_memory_kind(const memory &amemory) {
+    dnnl_sycl_interop_memory_kind_t ckind;
+    error::wrap_c_api(
+            dnnl_sycl_interop_memory_get_memory_kind(amemory.get(), &ckind),
+            "could not get memory kind");
+    return static_cast<memory_kind>(ckind);
+}
+
+#ifdef DNNL_EXPERIMENTAL_SPARSE
+/// Creates a memory object with multiple handles.
+///
+/// @param memory_desc Memory descriptor.
+/// @param aengine Engine to use.
+/// @param kind Memory allocation kind to specify the type of handles.
+/// @param handles Handles of the memory buffers to use as underlying storages.
+///     For each element of the @p handles array the following applies:
+///     - A USM pointer to the user-allocated buffer. In this case the library
+///       doesn't own the buffer. Requires @p memory_kind to be equal to
+///       dnnl::sycl_interop::memory_kind::usm.
+///     - A pointer to SYCL buffer. In this case the library doesn't own the
+///       buffer. Requires @p memory_kind be equal to be equal to
+///       dnnl::sycl_interop::memory_kind::buffer.
+///     - The DNNL_MEMORY_ALLOCATE special value. Instructs the library to
+///       allocate the buffer that corresponds to the memory allocation kind
+///       @p memory_kind for the memory object. In this case the library
+///       owns the buffer.
+///     - The DNNL_MEMORY_NONE specific value. Instructs the library to
+///       create memory object without an underlying buffer.
+///
+///  If the @p handles vector is not provided the library will allocate all
+///  buffers as if all handles have the special value DNNL_MEMORY_ALLOCATE.
+///
+/// @returns Created memory object.
+inline memory make_memory(const memory::desc &memory_desc,
+        const engine &aengine, memory_kind kind,
+        std::vector<void *> handles = {}) {
+    if (handles.empty()) {
+        const int nhandles = memory_desc.get_num_handles();
+        handles.resize(nhandles, DNNL_MEMORY_ALLOCATE);
+    }
+
+    dnnl_memory_t c_memory;
+    error::wrap_c_api(
+            dnnl_sycl_interop_memory_create_v2(&c_memory, memory_desc.get(),
+                    aengine.get(), convert_to_c(kind), (int)handles.size(),
+                    handles.data()),
+            "could not create a memory");
+    return memory(c_memory);
+}
+
+/// Creates a memory object.
+///
+/// Unless @p handle is equal to DNNL_MEMORY_NONE or DNNL_MEMORY_ALLOCATE, the
+/// constructed memory object will have the underlying buffer set. In this
+/// case, the buffer will be initialized as if:
+/// - dnnl::memory::set_data_handle() had been called, if @p memory_kind is
+///   equal to dnnl::sycl_interop::memory_kind::usm, or
+/// - dnnl::sycl_interop::set_buffer() has been called, if @p memory_kind is
+///   equal to dnnl::sycl_interop::memory_kind::buffer.
+///
+/// @param memory_desc Memory descriptor.
+/// @param aengine Engine to use.
+/// @param kind Memory allocation kind to specify the type of handle.
+/// @param handle Handle of the memory buffer to use as an underlying storage.
+///     - A USM pointer to the user-allocated buffer. In this case the library
+///       doesn't own the buffer. Requires @p memory_kind to be equal to
+///       dnnl::sycl_interop::memory_kind::usm.
+///     - A pointer to SYCL buffer. In this case the library doesn't own the
+///       buffer. Requires @p memory_kind be equal to be equal to
+///       dnnl::sycl_interop::memory_kind::buffer.
+///     - The DNNL_MEMORY_ALLOCATE special value. Instructs the library to
+///       allocate the buffer that corresponds to the memory allocation kind
+///       @p memory_kind for the memory object. In this case the library
+///       owns the buffer.
+///     - The DNNL_MEMORY_NONE specific value. Instructs the library to
+///       create memory object without an underlying buffer.
+///
+/// @returns Created memory object.
+inline memory make_memory(const memory::desc &memory_desc,
+        const engine &aengine, memory_kind kind, void *handle) {
+    return make_memory(
+            memory_desc, aengine, kind, std::vector<void *> {handle});
+}
+#else
+
+/// Creates a memory object.
+///
+/// Unless @p handle is equal to DNNL_MEMORY_NONE or DNNL_MEMORY_ALLOCATE, the
+/// constructed memory object will have the underlying buffer set. In this
+/// case, the buffer will be initialized as if:
+/// - dnnl::memory::set_data_handle() had been called, if @p memory_kind is
+///   equal to dnnl::sycl_interop::memory_kind::usm, or
+/// - dnnl::sycl_interop::set_buffer() has been called, if @p memory_kind is
+///   equal to dnnl::sycl_interop::memory_kind::buffer.
+///
+/// @param memory_desc Memory descriptor.
+/// @param aengine Engine to use.
+/// @param kind Memory allocation kind to specify the type of handle.
+/// @param handle Handle of the memory buffer to use as an underlying storage.
+///     - A USM pointer to the user-allocated buffer. In this case the library
+///       doesn't own the buffer. Requires @p memory_kind to be equal to
+///       dnnl::sycl_interop::memory_kind::usm.
+///     - A pointer to SYCL buffer. In this case the library doesn't own the
+///       buffer. Requires @p memory_kind be equal to be equal to
+///       dnnl::sycl_interop::memory_kind::buffer.
+///     - The DNNL_MEMORY_ALLOCATE special value. Instructs the library to
+///       allocate the buffer that corresponds to the memory allocation kind
+///       @p memory_kind for the memory object. In this case the library
+///       owns the buffer.
+///     - The DNNL_MEMORY_NONE specific value. Instructs the library to
+///       create memory object without an underlying buffer.
+///
+/// @returns Created memory object.
+inline memory make_memory(const memory::desc &memory_desc,
+        const engine &aengine, memory_kind kind,
+        void *handle = DNNL_MEMORY_ALLOCATE) {
+    dnnl_memory_t c_memory;
+    error::wrap_c_api(
+            dnnl_sycl_interop_memory_create(&c_memory, memory_desc.get(),
+                    aengine.get(), convert_to_c(kind), handle),
+            "could not create a memory");
+    return memory(c_memory);
+}
+#endif
+
+/// Constructs a memory object from a SYCL buffer.
+///
+/// @param memory_desc Memory descriptor.
+/// @param aengine Engine to use.
+/// @param abuffer A SYCL buffer to use.
+///
+/// @returns Created memory object.
+template <typename T, int ndims = 1>
+memory make_memory(const memory::desc &memory_desc, const engine &aengine,
+        sycl::buffer<T, ndims> &abuffer) {
+    memory amemory(memory_desc, aengine, DNNL_MEMORY_NONE);
+    set_buffer(amemory, abuffer);
+    return amemory;
+}
+
+/// Executes computations specified by the primitive in a specified stream and
+/// returns a SYCL event.
+///
+/// Arguments are passed via an arguments map containing
+/// <index, memory object> pairs. The index must be one of the `DNNL_ARG_*`
+/// values such as `DNNL_ARG_SRC`, and the memory must have a memory descriptor
+/// matching the one returned by
+/// #dnnl::primitive_desc::query_md(#query::exec_arg_md, index) unless using
+/// dynamic shapes (see #DNNL_RUNTIME_DIM_VAL).
+///
+/// @param aprimitive Primitive to execute.
+/// @param astream Stream object. The stream must belong to the same engine
+///     as the primitive.
+/// @param args Arguments map.
+/// @param deps Optional vector with `sycl::event` dependencies.
+///
+/// @returns Output event.
+inline sycl::event execute(const dnnl::primitive &aprimitive,
+        const stream &astream, const std::unordered_map<int, memory> &args,
+        const std::vector<sycl::event> &deps = {}) {
+    std::vector<dnnl_exec_arg_t> c_args;
+    c_args.reserve(args.size());
+    for (const auto &a : args)
+        c_args.push_back({a.first, a.second.get()});
+
+    sycl::event return_event;
+    error::wrap_c_api(
+            dnnl_sycl_interop_primitive_execute(aprimitive.get(), astream.get(),
+                    (int)c_args.size(), c_args.data(), &deps, &return_event),
+            "could not execute a primitive");
+    return return_event;
+}
+
+} // namespace sycl_interop
+
+/// @} dnnl_api_sycl_interop
+
+/// @} dnnl_api_interop
+
+} // namespace dnnl
+
+/// @} dnnl_api
+
+#endif // DNNL_SYCL_HPP

phivenv/Lib/site-packages/torch/include/oneapi/dnnl/dnnl_sycl_types.h

@@ -0,0 +1,51 @@
+/*******************************************************************************
+* Copyright 2020-2021 Intel Corporation
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*******************************************************************************/
+
+#ifndef ONEAPI_DNNL_DNNL_SYCL_TYPES_H
+#define ONEAPI_DNNL_DNNL_SYCL_TYPES_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/// @addtogroup dnnl_api
+/// @{
+
+/// @addtogroup dnnl_api_interop
+/// @{
+
+/// @addtogroup dnnl_api_sycl_interop
+/// @{
+
+/// Memory allocation kind.
+typedef enum {
+    /// USM (device, shared, host, or unknown) memory allocation kind - default.
+    dnnl_sycl_interop_usm,
+    /// Buffer memory allocation kind.
+    dnnl_sycl_interop_buffer,
+} dnnl_sycl_interop_memory_kind_t;
+
+/// @} dnnl_api_sycl_interop
+
+/// @} dnnl_api_interop
+
+/// @} dnnl_api
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif

phivenv/Lib/site-packages/torch/include/oneapi/dnnl/dnnl_threadpool.h

@@ -0,0 +1,118 @@
+/*******************************************************************************
+* Copyright 2020-2022 Intel Corporation
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*******************************************************************************/
+
+#ifndef ONEAPI_DNNL_DNNL_THREADPOOL_H
+#define ONEAPI_DNNL_DNNL_THREADPOOL_H
+
+#include "oneapi/dnnl/dnnl_config.h"
+#include "oneapi/dnnl/dnnl_types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/// @addtogroup dnnl_api
+/// @{
+
+/// @addtogroup dnnl_api_interop
+/// @{
+
+/// @addtogroup dnnl_api_threadpool_interop
+/// @{
+
+/// Creates an execution stream with specified threadpool.
+///
+/// @sa @ref dev_guide_threadpool
+///
+/// @param stream Output execution stream.
+/// @param engine Engine to create the execution stream on.
+/// @param threadpool Pointer to an instance of a C++ class that implements
+///     dnnl::threapdool_iface interface.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_threadpool_interop_stream_create(
+        dnnl_stream_t *stream, dnnl_engine_t engine, void *threadpool);
+
+/// Returns a threadpool to be used by the execution stream.
+///
+/// @sa @ref dev_guide_threadpool
+///
+/// @param astream Execution stream.
+/// @param threadpool Output pointer to an instance of a C++ class that
+///     implements dnnl::threapdool_iface interface. Set to NULL if the
+///     stream was created without threadpool.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_threadpool_interop_stream_get_threadpool(
+        dnnl_stream_t astream, void **threadpool);
+
+/// Sets the maximum concurrency assumed by oneDNN when outside a
+/// parallel call.
+///
+/// @param max_concurrency The maximum concurrency assumed by oneDNN
+/// when outside a parallel call. This is a threadlocal setting.
+/// @returns #dnnl_success on success and a status describing the
+/// error otherwise.
+dnnl_status_t DNNL_API dnnl_threadpool_interop_set_max_concurrency(
+        int max_concurrency);
+
+/// Gets the maximum concurrency assumed by oneDNN when outside a
+/// parallel call.
+///
+/// @param max_concurrency The maximum concurrency assumed by oneDNN
+/// when outside a parallel call. This is a threadlocal setting.
+/// @returns #dnnl_success on success and a status describing the
+/// error otherwise.
+dnnl_status_t DNNL_API dnnl_threadpool_interop_get_max_concurrency(
+        int *max_concurrency);
+
+/// @copydoc dnnl_sgemm()
+/// @param threadpool A pointer to a threadpool interface (only when built with
+///     the THREADPOOL CPU runtime).
+dnnl_status_t DNNL_API dnnl_threadpool_interop_sgemm(char transa, char transb,
+        dnnl_dim_t M, dnnl_dim_t N, dnnl_dim_t K, float alpha, const float *A,
+        dnnl_dim_t lda, const float *B, dnnl_dim_t ldb, float beta, float *C,
+        dnnl_dim_t ldc, void *threadpool);
+
+/// @copydoc dnnl_gemm_u8s8s32()
+/// @param threadpool A pointer to a threadpool interface (only when built with
+///     the THREADPOOL CPU runtime).
+dnnl_status_t DNNL_API dnnl_threadpool_interop_gemm_u8s8s32(char transa,
+        char transb, char offsetc, dnnl_dim_t M, dnnl_dim_t N, dnnl_dim_t K,
+        float alpha, const uint8_t *A, dnnl_dim_t lda, uint8_t ao,
+        const int8_t *B, dnnl_dim_t ldb, int8_t bo, float beta, int32_t *C,
+        dnnl_dim_t ldc, const int32_t *co, void *threadpool);
+
+/// @copydoc dnnl_gemm_s8s8s32()
+/// @param threadpool A pointer to a threadpool interface (only when built with
+///     the THREADPOOL CPU runtime).
+dnnl_status_t DNNL_API dnnl_threadpool_interop_gemm_s8s8s32(char transa,
+        char transb, char offsetc, dnnl_dim_t M, dnnl_dim_t N, dnnl_dim_t K,
+        float alpha, const int8_t *A, dnnl_dim_t lda, int8_t ao,
+        const int8_t *B, dnnl_dim_t ldb, int8_t bo, float beta, int32_t *C,
+        dnnl_dim_t ldc, const int32_t *co, void *threadpool);
+
+/// @} dnnl_api_threadpool_interop
+
+/// @} dnnl_api_interop
+
+/// @} dnnl_api
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif

phivenv/Lib/site-packages/torch/include/oneapi/dnnl/dnnl_threadpool.hpp

@@ -0,0 +1,113 @@
+/*******************************************************************************
+* Copyright 2020-2025 Intel Corporation
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*******************************************************************************/
+
+#ifndef ONEAPI_DNNL_DNNL_THREADPOOL_HPP
+#define ONEAPI_DNNL_DNNL_THREADPOOL_HPP
+
+#include "oneapi/dnnl/dnnl.hpp"
+#include "oneapi/dnnl/dnnl_threadpool.h"
+
+#include "oneapi/dnnl/dnnl_threadpool_iface.hpp"
+
+/// @addtogroup dnnl_api
+/// @{
+
+namespace dnnl {
+
+/// @addtogroup dnnl_api_interop
+/// @{
+
+/// @addtogroup dnnl_api_threadpool_interop Threadpool interoperability API
+/// API extensions to interact with the underlying Threadpool run-time.
+/// @{
+
+/// Threadpool interoperability namespace
+namespace threadpool_interop {
+
+/// Constructs an execution stream for the specified engine and threadpool.
+///
+/// @sa @ref dev_guide_threadpool
+///
+/// @param aengine Engine to create the stream on.
+/// @param threadpool Pointer to an instance of a C++ class that implements
+///     dnnl::threapdool_iface interface.
+/// @returns An execution stream.
+inline dnnl::stream make_stream(
+        const dnnl::engine &aengine, threadpool_iface *threadpool) {
+    dnnl_stream_t c_stream;
+    dnnl::error::wrap_c_api(dnnl_threadpool_interop_stream_create(
+                                    &c_stream, aengine.get(), threadpool),
+            "could not create stream");
+    return dnnl::stream(c_stream);
+}
+
+/// Returns the pointer to a threadpool that is used by an execution stream.
+///
+/// @sa @ref dev_guide_threadpool
+///
+/// @param astream An execution stream.
+/// @returns Output pointer to an instance of a C++ class that implements
+///     dnnl::threapdool_iface interface or NULL if the stream was created
+///     without threadpool.
+inline threadpool_iface *get_threadpool(const dnnl::stream &astream) {
+    void *tp;
+    dnnl::error::wrap_c_api(
+            dnnl_threadpool_interop_stream_get_threadpool(astream.get(), &tp),
+            "could not get stream threadpool");
+    return static_cast<threadpool_iface *>(tp);
+}
+
+/// @copydoc dnnl_threadpool_interop_sgemm()
+inline status sgemm(char transa, char transb, dnnl_dim_t M, dnnl_dim_t N,
+        dnnl_dim_t K, float alpha, const float *A, dnnl_dim_t lda,
+        const float *B, dnnl_dim_t ldb, float beta, float *C, dnnl_dim_t ldc,
+        threadpool_iface *threadpool) {
+    return static_cast<status>(dnnl_threadpool_interop_sgemm(transa, transb, M,
+            N, K, alpha, A, lda, B, ldb, beta, C, ldc, threadpool));
+}
+/// @copydoc dnnl_threadpool_interop_gemm_u8s8s32()
+inline status gemm_u8s8s32(char transa, char transb, char offsetc, dnnl_dim_t M,
+        dnnl_dim_t N, dnnl_dim_t K, float alpha, const uint8_t *A,
+        dnnl_dim_t lda, uint8_t ao, const int8_t *B, dnnl_dim_t ldb, int8_t bo,
+        float beta, int32_t *C, dnnl_dim_t ldc, const int32_t *co,
+        threadpool_iface *threadpool) {
+    return static_cast<status>(dnnl_threadpool_interop_gemm_u8s8s32(transa,
+            transb, offsetc, M, N, K, alpha, A, lda, ao, B, ldb, bo, beta, C,
+            ldc, co, threadpool));
+}
+
+/// @copydoc dnnl_threadpool_interop_gemm_s8s8s32()
+inline status gemm_s8s8s32(char transa, char transb, char offsetc, dnnl_dim_t M,
+        dnnl_dim_t N, dnnl_dim_t K, float alpha, const int8_t *A,
+        dnnl_dim_t lda, int8_t ao, const int8_t *B, dnnl_dim_t ldb, int8_t bo,
+        float beta, int32_t *C, dnnl_dim_t ldc, const int32_t *co,
+        threadpool_iface *threadpool) {
+    return static_cast<status>(dnnl_threadpool_interop_gemm_s8s8s32(transa,
+            transb, offsetc, M, N, K, alpha, A, lda, ao, B, ldb, bo, beta, C,
+            ldc, co, threadpool));
+}
+
+} // namespace threadpool_interop
+
+/// @} dnnl_api_threadpool_interop
+
+/// @} dnnl_api_interop
+
+} // namespace dnnl
+
+/// @} dnnl_api
+
+#endif

phivenv/Lib/site-packages/torch/include/oneapi/dnnl/dnnl_threadpool_iface.hpp

@@ -0,0 +1,73 @@
+/*******************************************************************************
+* Copyright 2020-2024 Intel Corporation
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*******************************************************************************/
+
+#ifndef ONEAPI_DNNL_DNNL_THREADPOOL_IFACE_HPP
+#define ONEAPI_DNNL_DNNL_THREADPOOL_IFACE_HPP
+
+#include <cstdint>
+#include <functional>
+
+/// @addtogroup dnnl_api
+/// @{
+
+namespace dnnl {
+
+/// @addtogroup dnnl_api_interop
+/// @{
+
+/// @addtogroup dnnl_api_threadpool_interop
+/// @{
+
+namespace threadpool_interop {
+
+/// Abstract threadpool interface. The users are expected to subclass this
+/// interface and pass an object to the library during CPU stream creation or
+/// directly in case of BLAS functions.
+struct threadpool_iface {
+    /// Returns the number of worker threads.
+    virtual int get_num_threads() const = 0;
+
+    /// Returns true if the calling thread belongs to this threadpool.
+    virtual bool get_in_parallel() const = 0;
+
+    /// Submits n instances of a closure for execution in parallel:
+    ///
+    /// for (int i = 0; i < n; i++) fn(i, n);
+    ///
+    virtual void parallel_for(int n, const std::function<void(int, int)> &fn)
+            = 0;
+
+    /// Returns threadpool behavior flags bit mask (see below).
+    virtual uint64_t get_flags() const = 0;
+
+    /// If set, parallel_for() returns immediately and oneDNN needs implement
+    /// waiting for the submitted closures to finish execution on its own.
+    static constexpr uint64_t ASYNCHRONOUS = 1;
+
+    virtual ~threadpool_iface() {}
+};
+
+} // namespace threadpool_interop
+
+/// @} dnnl_api_threadpool_interop
+
+/// @} dnnl_api_interop
+
+} // namespace dnnl
+
+/// @} dnnl_api
+
+#endif

phivenv/Lib/site-packages/torch/include/oneapi/dnnl/dnnl_ukernel.h

@@ -0,0 +1,337 @@
+/*******************************************************************************
+* Copyright 2024 Intel Corporation
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*******************************************************************************/
+
+/// @file
+/// ukernel C API
+
+#ifndef ONEAPI_DNNL_DNNL_UKERNEL_H
+#define ONEAPI_DNNL_DNNL_UKERNEL_H
+
+#include "oneapi/dnnl/dnnl.h"
+#include "oneapi/dnnl/dnnl_ukernel_types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/// @addtogroup dnnl_api
+/// @{
+
+/// @addtogroup dnnl_api_ukernel
+/// @{
+
+#ifdef DNNL_EXPERIMENTAL_UKERNEL
+
+/// Creates a ukernel attributes memory storage.
+///
+/// @param attr_params Output ukernel attributes memory storage.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_ukernel_attr_params_create(
+        dnnl_ukernel_attr_params_t *attr_params);
+
+/// Sets post-operations arguments to a storage.
+///
+/// @param attr_params Memory pointers storage object.
+/// @param post_ops_args A pointer to pointers of post_ops storages. Expected to
+///     be packed together.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_ukernel_attr_params_set_post_ops_args(
+        dnnl_ukernel_attr_params_t attr_params, const void **post_ops_args);
+
+/// Sets tensor A scales argument to a storage.
+///
+/// @param attr_params Memory pointers storage object.
+/// @param a_scales Pointer to the scales storage.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_ukernel_attr_params_set_A_scales(
+        dnnl_ukernel_attr_params_t attr_params, const void *a_scales);
+
+/// Sets tensor B scales argument to a storage.
+///
+/// If `dnnl_brgemm_set_B_scales` used mask of 2, then at least N values of
+/// selected data type are expected.
+///
+/// @param attr_params Memory pointers storage object.
+/// @param b_scales Pointer to the scales storage.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_ukernel_attr_params_set_B_scales(
+        dnnl_ukernel_attr_params_t attr_params, const void *b_scales);
+
+/// Sets tensor D scales argument to a storage.
+///
+/// @param attr_params Memory pointers storage object.
+/// @param d_scales Pointer to the scales storage.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_ukernel_attr_params_set_D_scales(
+        dnnl_ukernel_attr_params_t attr_params, const void *d_scales);
+
+/// Destroys a ukernel attributes memory storage.
+///
+/// @param attr_params Memory pointers storage object to destroy.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_ukernel_attr_params_destroy(
+        dnnl_ukernel_attr_params_t attr_params);
+
+/// @addtogroup dnnl_api_ukernel_brgemm
+/// @{
+
+/// Creates a BRGeMM ukernel object. Operates by the following formula:
+/// `C = [A x B]`.
+///
+/// @param brgemm Output BRGeMM ukernel object.
+/// @param M Dimension M of tensor A.
+/// @param N Dimension N of tensor B.
+/// @param K Dimension K of tensors A and B.
+/// @param batch_size Number of batches to process.
+/// @param lda Leading dimension of tensor A.
+/// @param ldb Leading dimension of tensor B.
+/// @param ldc Leading dimension of tensor C.
+/// @param a_dt Data type of tensor A.
+/// @param b_dt Data type of tensor B.
+/// @param c_dt Data type of tensor C. Must be dnnl_f32.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_brgemm_create(dnnl_brgemm_t *brgemm, dnnl_dim_t M,
+        dnnl_dim_t N, dnnl_dim_t K, dnnl_dim_t batch_size, dnnl_dim_t lda,
+        dnnl_dim_t ldb, dnnl_dim_t ldc, dnnl_data_type_t a_dt,
+        dnnl_data_type_t b_dt, dnnl_data_type_t c_dt);
+
+/// Sets adding an intermediate result to the output tensor C instead of
+/// writing: `C += [A x B]`.
+///
+/// @param brgemm BRGeMM ukernel object.
+/// @param add_C Value to indicate addition. Can be `0` to skip addition, and
+///     `1` to apply addition.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_brgemm_set_add_C(dnnl_brgemm_t brgemm, int add_C);
+
+/// Sets post-operations to a BRGeMM ukernel object: `D = post-operations(C)`.
+///
+/// Post-operations applies if one of the following holds:
+/// * Non-empty attributes are specified.
+/// * Output data type `d_dt` is different from accumulation data type `c_dt`.
+///
+/// If any of conditions happens, the final call of the accumulation chain
+/// must be `dnnl_brgemm_execute_postops`, and `dnnl_brgemm_execute`, otherwise.
+///
+/// @param brgemm BRGeMM ukernel object.
+/// @param ldd Leading dimension of tensor D.
+/// @param d_dt Data type of tensor D.
+/// @param post_ops Primitive post operations attribute to extend the kernel
+///     operations.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_brgemm_set_post_ops(dnnl_brgemm_t brgemm,
+        dnnl_dim_t ldd, dnnl_data_type_t d_dt, const_dnnl_post_ops_t post_ops);
+
+/// Sets tensor A scales mask to a BRGeMM ukernel object.
+///
+/// For quantization flavor tensor A scales apply to accumulation buffer once C
+/// is ready.
+///
+/// @param brgemm BRGeMM ukernel object.
+/// @param a_scale_mask Tensor A scale mask. Can be `0` only.
+dnnl_status_t DNNL_API dnnl_brgemm_set_A_scales(
+        dnnl_brgemm_t brgemm, int a_scale_mask);
+
+/// Sets tensor B scales mask to a BRGeMM ukernel object.
+///
+/// For quantization flavor tensor B scales apply to accumulation buffer once C
+/// is ready.
+///
+/// @param brgemm BRGeMM ukernel object.
+/// @param b_scale_mask Tensor B scale mask. Can be `0` and `2` only.
+dnnl_status_t DNNL_API dnnl_brgemm_set_B_scales(
+        dnnl_brgemm_t brgemm, int b_scale_mask);
+
+/// Sets tensor D scales mask to a BRGeMM ukernel object.
+///
+/// For quantization flavor tensor D scales apply after all post-ops are
+/// applied.
+///
+/// @param brgemm BRGeMM ukernel object.
+/// @param d_scale_mask Tensor D scale mask. Can be `0` only.
+dnnl_status_t DNNL_API dnnl_brgemm_set_D_scales(
+        dnnl_brgemm_t brgemm, int d_scale_mask);
+
+/// Finalizes initialization of a BRGeMM ukernel object.
+///
+/// This step is mandatory to query information from the object.
+///
+/// @param brgemm Output BRGeMM ukernel object.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_brgemm_finalize(dnnl_brgemm_t brgemm);
+
+/// Returns the packing type expected by a tensor B of a BRGeMM ukernel object.
+///
+/// @param brgemm BRGeMM ukernel object.
+/// @param pack_type Output packing type. Can be `dnnl_brgemm_no_pack` if
+///     packing is not expected, and `dnnl_brgemm_pack_32`, otherwise.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_brgemm_get_B_pack_type(
+        const_dnnl_brgemm_t brgemm, dnnl_pack_type_t *pack_type);
+
+/// Returns the size of a scratchpad memory needed for the BRGeMM ukernel
+/// object.
+///
+/// @param brgemm BRGeMM ukernel object.
+/// @param size Output size of a buffer required for the BRGeMM ukernel object.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_brgemm_get_scratchpad_size(
+        const_dnnl_brgemm_t brgemm, size_t *size);
+
+/// Returns the flag indicating when the call to `dnnl_brgemm_execute_postops`
+/// is valid.
+///
+/// @param brgemm BRGeMM ukernel object.
+/// @param valid The flag indicating if `dnnl_brgemm_execute_postops` is valid
+///     for a given ukernel object. `1` is for valid and `0`, otherwise.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_brgemm_is_execute_postops_valid(
+        const_dnnl_brgemm_t brgemm, int *valid);
+
+/// Initializes the hardware-specific context. If no initialization required,
+/// returns the success status.
+///
+/// @param brgemm BRGeMM ukernel object.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_brgemm_set_hw_context(const_dnnl_brgemm_t brgemm);
+
+/// Releases the hardware-specific context. Must be used after all the execution
+/// calls to BRGeMM ukernel objects.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_brgemm_release_hw_context();
+
+/// Generates an executable part of BRGeMM ukernel object.
+/// @param brgemm BRGeMM ukernel object.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_brgemm_generate(dnnl_brgemm_t brgemm);
+
+/// Executes a BRGeMM ukernel object.
+///
+/// @param brgemm BRGeMM ukernel object.
+/// @param A_ptr Base pointer to a tensor A.
+/// @param B_ptr Base pointer to a tensor B.
+/// @param A_B_offsets Pointer to the set of tensor A and tensor B offsets for
+///     each batch; the set must be contiguous in memory. Single batch should
+///     supply offsets for both tensors A and B simultaneously. The number of
+///     batches must coincide with the `batch_size` value passed at the creation
+///     stage.
+/// @param C_ptr Pointer to a tensor C (accumulation buffer).
+/// @param scratchpad_ptr Pointer to a scratchpad buffer.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_brgemm_execute(const_dnnl_brgemm_t brgemm,
+        const void *A_ptr, const void *B_ptr, const dnnl_dim_t *A_B_offsets,
+        void *C_ptr, void *scratchpad_ptr);
+
+/// Executes a BRGeMM ukernel object with post operations.
+///
+/// @param brgemm BRGeMM ukernel object.
+/// @param A Base pointer to a tensor A.
+/// @param B Base pointer to a tensor B.
+/// @param A_B_offsets Pointer to a set of tensor A and tensor B offsets for
+///     each batch. A set must be contiguous in memory. A single batch should
+///     supply offsets for both tensors A and B simultaneously. The number of
+///     batches must coincide with the `batch_size` value passed at the creation
+///     stage.
+/// @param C_ptr Pointer to a tensor C (accumulation buffer).
+/// @param D_ptr Pointer to a tensor D (output buffer).
+/// @param scratchpad_ptr Pointer to a scratchpad buffer.
+/// @param attr_params Ukernel attributes memory storage.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_brgemm_execute_postops(const_dnnl_brgemm_t brgemm,
+        const void *A, const void *B, const dnnl_dim_t *A_B_offsets,
+        const void *C_ptr, void *D_ptr, void *scratchpad_ptr,
+        const_dnnl_ukernel_attr_params_t attr_params);
+
+/// Destroys a BRGeMM ukernel object.
+///
+/// @param brgemm BRGeMM ukernel object to destroy.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_brgemm_destroy(dnnl_brgemm_t brgemm);
+
+/// Creates a transform object.
+///
+/// @param transform Output transform object.
+/// @param K Dimension K.
+/// @param N Dimension N.
+/// @param in_pack_type Input packing type. Must be one of
+///     `dnnl_pack_type_no_trans`, or `dnnl_pack_type_trans`.
+/// @param in_ld Input leading dimension.
+/// @param out_ld Output leading dimension. When packing data, it specifies a
+///     block by N dimension.
+/// @param in_dt Input data type.
+/// @param out_dt Output data type.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_transform_create(dnnl_transform_t *transform,
+        dnnl_dim_t K, dnnl_dim_t N, dnnl_pack_type_t in_pack_type,
+        dnnl_dim_t in_ld, dnnl_dim_t out_ld, dnnl_data_type_t in_dt,
+        dnnl_data_type_t out_dt);
+
+/// Generates an executable part of transform object.
+/// @param transform Transform object.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_transform_generate(dnnl_transform_t transform);
+
+/// Executes a transform object.
+///
+/// @param transform Transform object.
+/// @param in_ptr Pointer to an input buffer.
+/// @param out_ptr Pointer to an output buffer.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_transform_execute(
+        const_dnnl_transform_t transform, const void *in_ptr, void *out_ptr);
+
+/// Destroys a transform object.
+///
+/// @param transform Transform object.
+/// @returns #dnnl_success on success and a status describing the error
+///     otherwise.
+dnnl_status_t DNNL_API dnnl_transform_destroy(dnnl_transform_t transform);
+
+/// @} dnnl_api_ukernel_brgemm
+
+#endif
+
+/// @} dnnl_api_ukernel
+
+/// @} dnnl_api
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* ONEAPI_DNNL_DNNL_UKERNEL_H */

phivenv/Lib/site-packages/torch/include/oneapi/dnnl/dnnl_ukernel.hpp

@@ -0,0 +1,465 @@
+/*******************************************************************************
+* Copyright 2024-2025 Intel Corporation
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*******************************************************************************/
+
+/// @file
+/// ukernel C++ API
+
+#ifndef ONEAPI_DNNL_DNNL_UKERNEL_HPP
+#define ONEAPI_DNNL_DNNL_UKERNEL_HPP
+
+#include "oneapi/dnnl/dnnl.hpp"
+#include "oneapi/dnnl/dnnl_ukernel.h"
+
+/// @addtogroup dnnl_api oneDNN API
+/// @{
+
+/// oneDNN namespace
+namespace dnnl {
+
+#ifdef DNNL_EXPERIMENTAL_UKERNEL
+
+/// @addtogroup dnnl_api_utils
+/// @{
+
+/// @cond DO_NOT_DOCUMENT_THIS
+
+template <>
+struct handle_traits<dnnl_brgemm_t> {
+    static dnnl_status_t destructor(dnnl_brgemm_t p) {
+        return dnnl_brgemm_destroy(p);
+    }
+};
+
+template <>
+struct handle_traits<dnnl_transform_t> {
+    static dnnl_status_t destructor(dnnl_transform_t p) {
+        return dnnl_transform_destroy(p);
+    }
+};
+
+template <>
+struct handle_traits<dnnl_ukernel_attr_params_t> {
+    static dnnl_status_t destructor(dnnl_ukernel_attr_params_t p) {
+        return dnnl_ukernel_attr_params_destroy(p);
+    }
+};
+
+/// @endcond
+
+/// @} dnnl_api_utils
+
+#endif
+
+/// @addtogroup dnnl_api_ukernel Ukernels
+/// Collection of ukernels
+/// @{
+
+/// ukernel namespace
+namespace ukernel {
+
+#ifdef DNNL_EXPERIMENTAL_UKERNEL
+
+/// @addtogroup dnnl_api_ukernel_utils ukernel utils
+/// ukernel utility functions
+/// @{
+
+/// Packing specification
+enum class pack_type {
+    /// Undefined pack type. A guard value.
+    undef = dnnl_pack_type_undef,
+    /// Plain, not transposed layout. Similar to format_tag::ab.
+    no_trans = dnnl_pack_type_no_trans,
+    /// Plain, transposed layout. Similar to format_tag::ba.
+    trans = dnnl_pack_type_trans,
+    /// Packed by 32 bits along K dimension layout.
+    pack32 = dnnl_pack_type_pack32,
+};
+
+/// Ukernel attributes memory storage
+struct attr_params : public handle<dnnl_ukernel_attr_params_t> {
+    /// Constructs a ukernel attributes memory storage.
+    attr_params() {
+        dnnl_ukernel_attr_params_t c_params = nullptr;
+        dnnl_status_t status = dnnl_ukernel_attr_params_create(&c_params);
+        error::wrap_c_api(
+                status, "could not create an attributes memory storage");
+        reset(c_params);
+    }
+
+    /// Sets post-operations arguments to a storage.
+    ///
+    /// @param post_ops_args Pointer to pointers of post_ops storages.
+    ///     Expected to be packed together.
+    void set_post_ops_args(const void **post_ops_args) {
+        dnnl_status_t status = dnnl_ukernel_attr_params_set_post_ops_args(
+                get(), post_ops_args);
+        if (status != dnnl_success)
+            error::wrap_c_api(
+                    status, "could not set post operations arguments");
+    }
+
+    /// Sets tensor A scales arguments to a storage.
+    ///
+    /// @param a_scales Pointer to scales storage.
+    void set_A_scales(const void *a_scales) {
+        dnnl_status_t status
+                = dnnl_ukernel_attr_params_set_A_scales(get(), a_scales);
+        if (status != dnnl_success)
+            error::wrap_c_api(status, "could not set A scales argument");
+    }
+
+    /// Sets tensor B scales arguments to a storage.
+    ///
+    /// If @ref attr_params::set_B_scales used mask of 2, then at
+    /// least N values of selected data type are expected.
+    ///
+    /// @param b_scales Pointer to scales storage.
+    void set_B_scales(const void *b_scales) {
+        dnnl_status_t status
+                = dnnl_ukernel_attr_params_set_B_scales(get(), b_scales);
+        if (status != dnnl_success)
+            error::wrap_c_api(status, "could not set B scales argument");
+    }
+
+    /// Sets tensor D scales arguments to a storage.
+    ///
+    /// @param d_scales Pointer to scales storage.
+    void set_D_scales(const void *d_scales) {
+        dnnl_status_t status
+                = dnnl_ukernel_attr_params_set_D_scales(get(), d_scales);
+        if (status != dnnl_success)
+            error::wrap_c_api(status, "could not set D scales argument");
+    }
+};
+/// @} dnnl_api_ukernel_utils
+
+/// @addtogroup dnnl_api_ukernel_brgemm BRGeMM ukernel
+/// BRGeMM ukernel routines
+/// @{
+
+/// BRGeMM ukernel
+struct brgemm : public handle<dnnl_brgemm_t> {
+    /// Default constructor. Produces an empty object.
+    brgemm() = default;
+
+    /// Constructs a BRGeMM ukernel object. Operates by the following formula:
+    /// `C = [A x B]`.
+    ///
+    /// @param M Dimension M of tensor A.
+    /// @param N Dimension N of tensor B.
+    /// @param K Dimension K of tensors A and B.
+    /// @param batch_size Number of batches to process.
+    /// @param lda Leading dimension of tensor A.
+    /// @param ldb Leading dimension of tensor B.
+    /// @param ldc Leading dimension of tensor C.
+    /// @param a_dt Data type of tensor A.
+    /// @param b_dt Data type of tensor B.
+    /// @param c_dt Data type of tensor C.
+    /// @param allow_empty A flag signifying whether construction is
+    ///     allowed to fail without throwing an exception. In this case an
+    ///     empty object will be produced. This flag is optional and
+    ///     defaults to false.
+    brgemm(memory::dim M, memory::dim N, memory::dim K, memory::dim batch_size,
+            memory::dim lda, memory::dim ldb, memory::dim ldc,
+            memory::data_type a_dt, memory::data_type b_dt,
+            memory::data_type c_dt, bool allow_empty = false) {
+
+        dnnl_brgemm_t brgemm = nullptr;
+        dnnl_status_t status = dnnl_brgemm_create(&brgemm, M, N, K, batch_size,
+                lda, ldb, ldc, memory::convert_to_c(a_dt),
+                memory::convert_to_c(b_dt), memory::convert_to_c(c_dt));
+
+        if (!allow_empty)
+            error::wrap_c_api(
+                    status, "could not create a BRGeMM ukernel object");
+        reset(brgemm);
+    }
+
+    /// Sets adding an intermediate result to the output tensor C instead of
+    /// writing: `C += [A x B]`.
+    ///
+    /// @param add_C Value to indicate addition. `false` to skip addition, and
+    ///     `true` to apply addition.
+    void set_add_C(bool add_C) {
+        dnnl_status_t status
+                = dnnl_brgemm_set_add_C(get(), static_cast<int>(add_C));
+        if (status != dnnl_success)
+            error::wrap_c_api(status, "could not set add_C attribute");
+    }
+
+    /// Sets post-operations to a BRGeMM ukernel object:
+    /// `D = post-operations(C)`.
+    ///
+    /// Post-operations applies if one of the following holds:
+    /// * Non-empty post-operations are specified.
+    /// * Output data type `d_dt` is different from accumulation data type
+    ///     `c_dt`.
+    ///
+    /// @param ldd Leading dimension of tensor D.
+    /// @param d_dt Data type of tensor D.
+    /// @param po Primitive post-operation attributes to extend the kernel
+    ///     operations.
+    void set_post_ops(memory::dim ldd, memory::data_type d_dt,
+            const post_ops &po = default_post_ops()) {
+        dnnl_status_t status = dnnl_brgemm_set_post_ops(
+                get(), ldd, memory::convert_to_c(d_dt), po.get());
+        if (status != dnnl_success)
+            error::wrap_c_api(status, "could not set post operations");
+    }
+
+    /// Sets tensor A scales mask to a BRGeMM ukernel object.
+    ///
+    /// For quantization flavor tensor A scales apply to accumulation buffer
+    /// once C is ready.
+    ///
+    /// @param a_scale_mask Tensor A scale mask. Can be `0` only.
+    void set_A_scales(int a_scale_mask) {
+        dnnl_status_t status = dnnl_brgemm_set_A_scales(get(), a_scale_mask);
+        if (status != dnnl_success)
+            error::wrap_c_api(status, "could not set A scales");
+    }
+
+    /// Sets tensor B scales mask to a BRGeMM ukernel object.
+    ///
+    /// For quantization flavor tensor B scales apply to accumulation buffer
+    /// once C is ready.
+    ///
+    /// @param b_scale_mask Tensor B scale mask. Can be `0` and `2` only.
+    void set_B_scales(int b_scale_mask) {
+        dnnl_status_t status = dnnl_brgemm_set_B_scales(get(), b_scale_mask);
+        if (status != dnnl_success)
+            error::wrap_c_api(status, "could not set B scales");
+    }
+
+    /// Sets tensor D scales mask to a BRGeMM ukernel object.
+    ///
+    /// For quantization flavor tensor D scales apply after all post-ops are
+    /// applied.
+    ///
+    /// @param d_scale_mask Tensor D scale mask. Can be `0` only.
+    void set_D_scales(int d_scale_mask) {
+        dnnl_status_t status = dnnl_brgemm_set_D_scales(get(), d_scale_mask);
+        if (status != dnnl_success)
+            error::wrap_c_api(status, "could not set D scales");
+    }
+
+    /// Finalizes initialization of a BRGeMM ukernel object.
+    ///
+    /// This step must be performed prior to querying information from the
+    /// object.
+    void finalize() {
+        dnnl_status_t status = dnnl_brgemm_finalize(get());
+        if (status != dnnl_success)
+            error::wrap_c_api(status, "could not finalize an object");
+    }
+
+    /// Returns the packing type expected by a tensor B of a BRGeMM ukernel
+    /// object.
+    pack_type get_B_pack_type() const {
+        dnnl_pack_type_t c_pack_type;
+        dnnl_status_t status = dnnl_brgemm_get_B_pack_type(get(), &c_pack_type);
+        if (status != dnnl_success)
+            error::wrap_c_api(status, "could not query B pack type");
+
+        return static_cast<pack_type>(c_pack_type);
+    }
+
+    /// Returns the size of a scratchpad memory needed for the BRGeMM ukernel
+    /// object.
+    size_t get_scratchpad_size() const {
+        size_t size;
+        dnnl_status_t status = dnnl_brgemm_get_scratchpad_size(get(), &size);
+        if (status != dnnl_success)
+            error::wrap_c_api(status,
+                    "could not query a scratchpad size from a BRGeMM ukernel "
+                    "object");
+        return size;
+    }
+
+    /// Returns the flag indicating when the call to execute with post
+    /// operations is valid.
+    ///
+    /// `True` is for a valid call, `false`, otherwise.
+    bool is_execute_postops_valid() const {
+        int valid;
+        dnnl_status_t status
+                = dnnl_brgemm_is_execute_postops_valid(get(), &valid);
+        if (status != dnnl_success)
+            error::wrap_c_api(status,
+                    "could not query a flag for execute postops from a BRGeMM "
+                    "ukernel object");
+        return static_cast<bool>(valid);
+    }
+
+    /// Initializes the hardware-specific context. Affects the global state for
+    /// all BRGeMM ukernel objects. If no initialization required, returns.
+    void set_hw_context() const {
+        dnnl_status_t status = dnnl_brgemm_set_hw_context(get());
+        if (status != dnnl_success)
+            error::wrap_c_api(status, "could not set hardware context");
+    }
+
+    /// Releases the hardware-specific context. Affects the global state for
+    /// all BRGeMM ukernel objects. Must be used after all the execution calls
+    /// to BRGeMM ukernel objects.
+    static void release_hw_context() {
+        dnnl_status_t status = dnnl_brgemm_release_hw_context();
+        if (status != dnnl_success)
+            error::wrap_c_api(status, "could not release hardware context");
+    }
+
+    /// Generates an executable part of BRGeMM ukernel object.
+    void generate() {
+        dnnl_status_t status = dnnl_brgemm_generate(get());
+        if (status != dnnl_success)
+            error::wrap_c_api(status, "could not generate a kernel");
+    }
+
+    /// Executes a BRGeMM ukernel object.
+    ///
+    /// @param A Base pointer to a tensor A.
+    /// @param B Base pointer to a tensor B.
+    /// @param A_B_offsets Vector of pairs of tensors A and B offsets for
+    ///     each batch. The number of batches must coincide with the
+    ///     `batch_size` value passed at object construction stage.
+    /// @param C Pointer to a tensor C (accumulation buffer).
+    /// @param scratchpad Pointer to a scratchpad buffer.
+    void execute(const void *A, const void *B,
+            const std::vector<std::pair<memory::dim, memory::dim>> &A_B_offsets,
+            void *C, void *scratchpad) const {
+        // TODO: export batch_element to C API later for user to fill it and
+        // pass directly to the call.
+        dnnl_status_t status = dnnl_brgemm_execute(get(), A, B,
+                (const dnnl_dim_t *)A_B_offsets.data(), C, scratchpad);
+        if (status != dnnl_success)
+            error::wrap_c_api(
+                    status, "could not execute a BRGeMM ukernel object");
+    }
+
+    /// Executes a BRGeMM ukernel object with post operations.
+    ///
+    /// @param A Base pointer to a tensor A.
+    /// @param B Base pointer to a tensor B.
+    /// @param A_B_offsets Vector of pairs of tensors A and B offsets for
+    ///     each batch. The number of batches must coincide with the
+    ///     `batch_size` value passed at object construction stage.
+    /// @param C Pointer to a tensor C (accumulation buffer).
+    /// @param D Pointer to a tensor D (output buffer).
+    /// @param scratchpad Pointer to a scratchpad buffer.
+    /// @param params Post-op memory arguments. Must be passed If binary
+    ///     post-op or scales were set.
+    void execute(const void *A, const void *B,
+            const std::vector<std::pair<memory::dim, memory::dim>> &A_B_offsets,
+            const void *C, void *D, void *scratchpad,
+            const attr_params &params = default_attr_params()) const {
+        // TODO: export batch_element to C API later for user to fill it and
+        // pass directly to the call.
+        dnnl_status_t status = dnnl_brgemm_execute_postops(get(), A, B,
+                (const dnnl_dim_t *)A_B_offsets.data(), C, D, scratchpad,
+                params.get());
+        if (status != dnnl_success)
+            error::wrap_c_api(
+                    status, "could not execute a BRGeMM ukernel object");
+    }
+
+    /// Returns a constant reference to a static instance of default constructed
+    /// primitive post-operations attribute.
+    static const post_ops &default_post_ops() {
+        static const post_ops po;
+        return po;
+    }
+
+    /// Returns a constant reference to a static instance of default constructed
+    /// ukernel attributes parameters.
+    static const attr_params &default_attr_params() {
+        static const attr_params ap;
+        return ap;
+    }
+};
+/// @} dnnl_api_ukernel_brgemm
+
+/// @addtogroup dnnl_api_ukernel_transform Transform ukernel
+/// Transform routines
+/// @{
+
+/// Transform ukernel
+struct transform : public handle<dnnl_transform_t> {
+    /// Default constructor. Produces an empty object.
+    transform() = default;
+
+    /// Constructs a transform object.
+    ///
+    /// @param K Dimension K.
+    /// @param N Dimension N.
+    /// @param in_pack_type Input packing type. Must be one of
+    ///     `pack_type::no_trans`, or `pack_type::trans`.
+    /// @param in_ld Input leading dimension.
+    /// @param out_ld Output leading dimension. Specifies a block by N dimension
+    ///     during data packing.
+    /// @param in_dt Input data type.
+    /// @param out_dt Output data type.
+    /// @param allow_empty A flag signifying whether construction is
+    ///     allowed to fail without throwing an exception. In this case an
+    ///     empty object will be produced. This flag is optional and
+    ///     defaults to false.
+    transform(memory::dim K, memory::dim N, pack_type in_pack_type,
+            memory::dim in_ld, memory::dim out_ld, memory::data_type in_dt,
+            memory::data_type out_dt, bool allow_empty = false) {
+
+        dnnl_transform_t transform = nullptr;
+        dnnl_status_t status = dnnl_transform_create(&transform, K, N,
+                static_cast<dnnl_pack_type_t>(in_pack_type), in_ld, out_ld,
+                memory::convert_to_c(in_dt), memory::convert_to_c(out_dt));
+
+        if (!allow_empty)
+            error::wrap_c_api(status,
+                    "could not create a BRGeMM ukernel packing B object");
+        reset(transform);
+    }
+
+    /// Generates an executable part of transform object.
+    void generate() {
+        dnnl_status_t status = dnnl_transform_generate(get());
+        if (status != dnnl_success)
+            error::wrap_c_api(status,
+                    "could not generate a BRGeMM ukernel packing B object");
+    }
+
+    /// Executes a transform object.
+    ///
+    /// @param in Pointer to an input buffer.
+    /// @param out Pointer to an output buffer.
+    void execute(const void *in, void *out) const {
+        dnnl_status_t status = dnnl_transform_execute(get(), in, out);
+        if (status != dnnl_success)
+            error::wrap_c_api(status,
+                    "could not execute a BRGeMM ukernel packing B object");
+    }
+};
+
+/// @} dnnl_api_ukernel_transform
+
+#endif
+
+} // namespace ukernel
+
+/// @} dnnl_api_ukernel
+
+} // namespace dnnl
+
+/// @} dnnl_api
+
+#endif /* ONEAPI_DNNL_DNNL_UKERNEL_HPP */

phivenv/Lib/site-packages/torch/include/oneapi/dnnl/dnnl_ukernel_types.h

@@ -0,0 +1,93 @@
+/*******************************************************************************
+* Copyright 2024 Intel Corporation
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*******************************************************************************/
+
+/// @file
+/// ukernel C API types definitions
+
+#ifndef ONEAPI_DNNL_DNNL_UKERNEL_TYPES_H
+#define ONEAPI_DNNL_DNNL_UKERNEL_TYPES_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#include "oneapi/dnnl/dnnl_types.h"
+
+/// @addtogroup dnnl_api
+/// @{
+
+/// @addtogroup dnnl_api_ukernel
+/// @{
+
+#ifdef DNNL_EXPERIMENTAL_UKERNEL
+
+/// Packing specification
+typedef enum {
+    /// Undefined pack type. A guard value.
+    dnnl_pack_type_undef = 0,
+    /// Plain, not transposed layout. Similar to format_tag::ab.
+    dnnl_pack_type_no_trans,
+    /// Plain, transposed layout. Similar to format_tag::ba.
+    dnnl_pack_type_trans,
+    /// Packed by 32 bits along K dimension layout.
+    dnnl_pack_type_pack32,
+} dnnl_pack_type_t;
+
+/// @struct dnnl_ukernel_attr_params
+/// An opaque structure to describe ukernel attributes memory storage.
+struct dnnl_ukernel_attr_params;
+
+/// A ukernel attributes memory storage handle.
+typedef struct dnnl_ukernel_attr_params *dnnl_ukernel_attr_params_t;
+
+/// A constant ukernel attributes memory storage handle.
+typedef const struct dnnl_ukernel_attr_params *const_dnnl_ukernel_attr_params_t;
+
+/// @addtogroup dnnl_api_ukernel_brgemm
+/// @{
+
+/// @struct dnnl_brgemm
+/// An opaque structure to describe a brgemm ukernel.
+struct dnnl_brgemm;
+
+/// A brgemm ukernel handle.
+typedef struct dnnl_brgemm *dnnl_brgemm_t;
+
+/// A constant brgemm ukernel handle.
+typedef const struct dnnl_brgemm *const_dnnl_brgemm_t;
+
+/// @struct dnnl_transform
+/// An opaque structure to describe a transform routine.
+struct dnnl_transform;
+
+/// A transform routine handle.
+typedef struct dnnl_transform *dnnl_transform_t;
+
+/// A constant transform routine handle.
+typedef const struct dnnl_transform *const_dnnl_transform_t;
+
+/// @} dnnl_api_ukernel_brgemm
+#endif
+
+/// @} dnnl_api_ukernel
+
+/// @} dnnl_api
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* ONEAPI_DNNL_DNNL_UKERNEL_TYPES_H */

phivenv/Lib/site-packages/torch/include/oneapi/dnnl/dnnl_version.h

@@ -0,0 +1,33 @@
+/*******************************************************************************
+* Copyright 2019-2024 Intel Corporation
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*******************************************************************************/
+
+#ifndef ONEAPI_DNNL_DNNL_VERSION_H
+#define ONEAPI_DNNL_DNNL_VERSION_H
+
+// clang-format off
+
+/// Major version
+#define DNNL_VERSION_MAJOR 3
+
+/// Minor version
+#define DNNL_VERSION_MINOR 7
+
+/// Patch version
+#define DNNL_VERSION_PATCH 1
+
+// clang-format on
+
+#endif

phivenv/Lib/site-packages/torch/include/oneapi/dnnl/dnnl_version_hash.h

@@ -0,0 +1,31 @@
+/*******************************************************************************
+* Copyright 2024 Intel Corporation
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+*     http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*******************************************************************************/
+
+#ifndef ONEAPI_DNNL_DNNL_VERSION_HASH_H
+#define ONEAPI_DNNL_DNNL_VERSION_HASH_H
+
+// clang-format off
+
+/// Note: this macro and header file were moved to a separate instance to avoid
+/// incremental build issues as moving from commit to commit would trigger a
+/// complete library rebuild. Including a generated header file in a single
+/// translation unit makes this problem go away.
+/// Git commit hash
+#define DNNL_VERSION_HASH  "8d263e693366ef8db40acc569cc7d8edf644556d"
+
+// clang-format on
+
+#endif

phivenv/Lib/site-packages/torch/include/pybind11/attr.h

@@ -0,0 +1,690 @@
+/*
+    pybind11/attr.h: Infrastructure for processing custom
+    type and function attributes
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "detail/common.h"
+#include "cast.h"
+
+#include <functional>
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+/// \addtogroup annotations
+/// @{
+
+/// Annotation for methods
+struct is_method {
+    handle class_;
+    explicit is_method(const handle &c) : class_(c) {}
+};
+
+/// Annotation for setters
+struct is_setter {};
+
+/// Annotation for operators
+struct is_operator {};
+
+/// Annotation for classes that cannot be subclassed
+struct is_final {};
+
+/// Annotation for parent scope
+struct scope {
+    handle value;
+    explicit scope(const handle &s) : value(s) {}
+};
+
+/// Annotation for documentation
+struct doc {
+    const char *value;
+    explicit doc(const char *value) : value(value) {}
+};
+
+/// Annotation for function names
+struct name {
+    const char *value;
+    explicit name(const char *value) : value(value) {}
+};
+
+/// Annotation indicating that a function is an overload associated with a given "sibling"
+struct sibling {
+    handle value;
+    explicit sibling(const handle &value) : value(value.ptr()) {}
+};
+
+/// Annotation indicating that a class derives from another given type
+template <typename T>
+struct base {
+
+    PYBIND11_DEPRECATED(
+        "base<T>() was deprecated in favor of specifying 'T' as a template argument to class_")
+    base() = default;
+};
+
+/// Keep patient alive while nurse lives
+template <size_t Nurse, size_t Patient>
+struct keep_alive {};
+
+/// Annotation indicating that a class is involved in a multiple inheritance relationship
+struct multiple_inheritance {};
+
+/// Annotation which enables dynamic attributes, i.e. adds `__dict__` to a class
+struct dynamic_attr {};
+
+/// Annotation which enables the buffer protocol for a type
+struct buffer_protocol {};
+
+/// Annotation which requests that a special metaclass is created for a type
+struct metaclass {
+    handle value;
+
+    PYBIND11_DEPRECATED("py::metaclass() is no longer required. It's turned on by default now.")
+    metaclass() = default;
+
+    /// Override pybind11's default metaclass
+    explicit metaclass(handle value) : value(value) {}
+};
+
+/// Specifies a custom callback with signature `void (PyHeapTypeObject*)` that
+/// may be used to customize the Python type.
+///
+/// The callback is invoked immediately before `PyType_Ready`.
+///
+/// Note: This is an advanced interface, and uses of it may require changes to
+/// work with later versions of pybind11.  You may wish to consult the
+/// implementation of `make_new_python_type` in `detail/classes.h` to understand
+/// the context in which the callback will be run.
+struct custom_type_setup {
+    using callback = std::function<void(PyHeapTypeObject *heap_type)>;
+
+    explicit custom_type_setup(callback value) : value(std::move(value)) {}
+
+    callback value;
+};
+
+/// Annotation that marks a class as local to the module:
+struct module_local {
+    const bool value;
+    constexpr explicit module_local(bool v = true) : value(v) {}
+};
+
+/// Annotation to mark enums as an arithmetic type
+struct arithmetic {};
+
+/// Mark a function for addition at the beginning of the existing overload chain instead of the end
+struct prepend {};
+
+/** \rst
+    A call policy which places one or more guard variables (``Ts...``) around the function call.
+
+    For example, this definition:
+
+    .. code-block:: cpp
+
+        m.def("foo", foo, py::call_guard<T>());
+
+    is equivalent to the following pseudocode:
+
+    .. code-block:: cpp
+
+        m.def("foo", [](args...) {
+            T scope_guard;
+            return foo(args...); // forwarded arguments
+        });
+ \endrst */
+template <typename... Ts>
+struct call_guard;
+
+template <>
+struct call_guard<> {
+    using type = detail::void_type;
+};
+
+template <typename T>
+struct call_guard<T> {
+    static_assert(std::is_default_constructible<T>::value,
+                  "The guard type must be default constructible");
+
+    using type = T;
+};
+
+template <typename T, typename... Ts>
+struct call_guard<T, Ts...> {
+    struct type {
+        T guard{}; // Compose multiple guard types with left-to-right default-constructor order
+        typename call_guard<Ts...>::type next{};
+    };
+};
+
+/// @} annotations
+
+PYBIND11_NAMESPACE_BEGIN(detail)
+/* Forward declarations */
+enum op_id : int;
+enum op_type : int;
+struct undefined_t;
+template <op_id id, op_type ot, typename L = undefined_t, typename R = undefined_t>
+struct op_;
+void keep_alive_impl(size_t Nurse, size_t Patient, function_call &call, handle ret);
+
+/// Internal data structure which holds metadata about a keyword argument
+struct argument_record {
+    const char *name;  ///< Argument name
+    const char *descr; ///< Human-readable version of the argument value
+    handle value;      ///< Associated Python object
+    bool convert : 1;  ///< True if the argument is allowed to convert when loading
+    bool none : 1;     ///< True if None is allowed when loading
+
+    argument_record(const char *name, const char *descr, handle value, bool convert, bool none)
+        : name(name), descr(descr), value(value), convert(convert), none(none) {}
+};
+
+/// Internal data structure which holds metadata about a bound function (signature, overloads,
+/// etc.)
+struct function_record {
+    function_record()
+        : is_constructor(false), is_new_style_constructor(false), is_stateless(false),
+          is_operator(false), is_method(false), is_setter(false), has_args(false),
+          has_kwargs(false), prepend(false) {}
+
+    /// Function name
+    char *name = nullptr; /* why no C++ strings? They generate heavier code.. */
+
+    // User-specified documentation string
+    char *doc = nullptr;
+
+    /// Human-readable version of the function signature
+    char *signature = nullptr;
+
+    /// List of registered keyword arguments
+    std::vector<argument_record> args;
+
+    /// Pointer to lambda function which converts arguments and performs the actual call
+    handle (*impl)(function_call &) = nullptr;
+
+    /// Storage for the wrapped function pointer and captured data, if any
+    void *data[3] = {};
+
+    /// Pointer to custom destructor for 'data' (if needed)
+    void (*free_data)(function_record *ptr) = nullptr;
+
+    /// Return value policy associated with this function
+    return_value_policy policy = return_value_policy::automatic;
+
+    /// True if name == '__init__'
+    bool is_constructor : 1;
+
+    /// True if this is a new-style `__init__` defined in `detail/init.h`
+    bool is_new_style_constructor : 1;
+
+    /// True if this is a stateless function pointer
+    bool is_stateless : 1;
+
+    /// True if this is an operator (__add__), etc.
+    bool is_operator : 1;
+
+    /// True if this is a method
+    bool is_method : 1;
+
+    /// True if this is a setter
+    bool is_setter : 1;
+
+    /// True if the function has a '*args' argument
+    bool has_args : 1;
+
+    /// True if the function has a '**kwargs' argument
+    bool has_kwargs : 1;
+
+    /// True if this function is to be inserted at the beginning of the overload resolution chain
+    bool prepend : 1;
+
+    /// Number of arguments (including py::args and/or py::kwargs, if present)
+    std::uint16_t nargs;
+
+    /// Number of leading positional arguments, which are terminated by a py::args or py::kwargs
+    /// argument or by a py::kw_only annotation.
+    std::uint16_t nargs_pos = 0;
+
+    /// Number of leading arguments (counted in `nargs`) that are positional-only
+    std::uint16_t nargs_pos_only = 0;
+
+    /// Python method object
+    PyMethodDef *def = nullptr;
+
+    /// Python handle to the parent scope (a class or a module)
+    handle scope;
+
+    /// Python handle to the sibling function representing an overload chain
+    handle sibling;
+
+    /// Pointer to next overload
+    function_record *next = nullptr;
+};
+
+/// Special data structure which (temporarily) holds metadata about a bound class
+struct type_record {
+    PYBIND11_NOINLINE type_record()
+        : multiple_inheritance(false), dynamic_attr(false), buffer_protocol(false),
+          default_holder(true), module_local(false), is_final(false) {}
+
+    /// Handle to the parent scope
+    handle scope;
+
+    /// Name of the class
+    const char *name = nullptr;
+
+    // Pointer to RTTI type_info data structure
+    const std::type_info *type = nullptr;
+
+    /// How large is the underlying C++ type?
+    size_t type_size = 0;
+
+    /// What is the alignment of the underlying C++ type?
+    size_t type_align = 0;
+
+    /// How large is the type's holder?
+    size_t holder_size = 0;
+
+    /// The global operator new can be overridden with a class-specific variant
+    void *(*operator_new)(size_t) = nullptr;
+
+    /// Function pointer to class_<..>::init_instance
+    void (*init_instance)(instance *, const void *) = nullptr;
+
+    /// Function pointer to class_<..>::dealloc
+    void (*dealloc)(detail::value_and_holder &) = nullptr;
+
+    /// List of base classes of the newly created type
+    list bases;
+
+    /// Optional docstring
+    const char *doc = nullptr;
+
+    /// Custom metaclass (optional)
+    handle metaclass;
+
+    /// Custom type setup.
+    custom_type_setup::callback custom_type_setup_callback;
+
+    /// Multiple inheritance marker
+    bool multiple_inheritance : 1;
+
+    /// Does the class manage a __dict__?
+    bool dynamic_attr : 1;
+
+    /// Does the class implement the buffer protocol?
+    bool buffer_protocol : 1;
+
+    /// Is the default (unique_ptr) holder type used?
+    bool default_holder : 1;
+
+    /// Is the class definition local to the module shared object?
+    bool module_local : 1;
+
+    /// Is the class inheritable from python classes?
+    bool is_final : 1;
+
+    PYBIND11_NOINLINE void add_base(const std::type_info &base, void *(*caster)(void *) ) {
+        auto *base_info = detail::get_type_info(base, false);
+        if (!base_info) {
+            std::string tname(base.name());
+            detail::clean_type_id(tname);
+            pybind11_fail("generic_type: type \"" + std::string(name)
+                          + "\" referenced unknown base type \"" + tname + "\"");
+        }
+
+        if (default_holder != base_info->default_holder) {
+            std::string tname(base.name());
+            detail::clean_type_id(tname);
+            pybind11_fail("generic_type: type \"" + std::string(name) + "\" "
+                          + (default_holder ? "does not have" : "has")
+                          + " a non-default holder type while its base \"" + tname + "\" "
+                          + (base_info->default_holder ? "does not" : "does"));
+        }
+
+        bases.append((PyObject *) base_info->type);
+
+#if PY_VERSION_HEX < 0x030B0000
+        dynamic_attr |= base_info->type->tp_dictoffset != 0;
+#else
+        dynamic_attr |= (base_info->type->tp_flags & Py_TPFLAGS_MANAGED_DICT) != 0;
+#endif
+
+        if (caster) {
+            base_info->implicit_casts.emplace_back(type, caster);
+        }
+    }
+};
+
+inline function_call::function_call(const function_record &f, handle p) : func(f), parent(p) {
+    args.reserve(f.nargs);
+    args_convert.reserve(f.nargs);
+}
+
+/// Tag for a new-style `__init__` defined in `detail/init.h`
+struct is_new_style_constructor {};
+
+/**
+ * Partial template specializations to process custom attributes provided to
+ * cpp_function_ and class_. These are either used to initialize the respective
+ * fields in the type_record and function_record data structures or executed at
+ * runtime to deal with custom call policies (e.g. keep_alive).
+ */
+template <typename T, typename SFINAE = void>
+struct process_attribute;
+
+template <typename T>
+struct process_attribute_default {
+    /// Default implementation: do nothing
+    static void init(const T &, function_record *) {}
+    static void init(const T &, type_record *) {}
+    static void precall(function_call &) {}
+    static void postcall(function_call &, handle) {}
+};
+
+/// Process an attribute specifying the function's name
+template <>
+struct process_attribute<name> : process_attribute_default<name> {
+    static void init(const name &n, function_record *r) { r->name = const_cast<char *>(n.value); }
+};
+
+/// Process an attribute specifying the function's docstring
+template <>
+struct process_attribute<doc> : process_attribute_default<doc> {
+    static void init(const doc &n, function_record *r) { r->doc = const_cast<char *>(n.value); }
+};
+
+/// Process an attribute specifying the function's docstring (provided as a C-style string)
+template <>
+struct process_attribute<const char *> : process_attribute_default<const char *> {
+    static void init(const char *d, function_record *r) { r->doc = const_cast<char *>(d); }
+    static void init(const char *d, type_record *r) { r->doc = d; }
+};
+template <>
+struct process_attribute<char *> : process_attribute<const char *> {};
+
+/// Process an attribute indicating the function's return value policy
+template <>
+struct process_attribute<return_value_policy> : process_attribute_default<return_value_policy> {
+    static void init(const return_value_policy &p, function_record *r) { r->policy = p; }
+};
+
+/// Process an attribute which indicates that this is an overloaded function associated with a
+/// given sibling
+template <>
+struct process_attribute<sibling> : process_attribute_default<sibling> {
+    static void init(const sibling &s, function_record *r) { r->sibling = s.value; }
+};
+
+/// Process an attribute which indicates that this function is a method
+template <>
+struct process_attribute<is_method> : process_attribute_default<is_method> {
+    static void init(const is_method &s, function_record *r) {
+        r->is_method = true;
+        r->scope = s.class_;
+    }
+};
+
+/// Process an attribute which indicates that this function is a setter
+template <>
+struct process_attribute<is_setter> : process_attribute_default<is_setter> {
+    static void init(const is_setter &, function_record *r) { r->is_setter = true; }
+};
+
+/// Process an attribute which indicates the parent scope of a method
+template <>
+struct process_attribute<scope> : process_attribute_default<scope> {
+    static void init(const scope &s, function_record *r) { r->scope = s.value; }
+};
+
+/// Process an attribute which indicates that this function is an operator
+template <>
+struct process_attribute<is_operator> : process_attribute_default<is_operator> {
+    static void init(const is_operator &, function_record *r) { r->is_operator = true; }
+};
+
+template <>
+struct process_attribute<is_new_style_constructor>
+    : process_attribute_default<is_new_style_constructor> {
+    static void init(const is_new_style_constructor &, function_record *r) {
+        r->is_new_style_constructor = true;
+    }
+};
+
+inline void check_kw_only_arg(const arg &a, function_record *r) {
+    if (r->args.size() > r->nargs_pos && (!a.name || a.name[0] == '\0')) {
+        pybind11_fail("arg(): cannot specify an unnamed argument after a kw_only() annotation or "
+                      "args() argument");
+    }
+}
+
+inline void append_self_arg_if_needed(function_record *r) {
+    if (r->is_method && r->args.empty()) {
+        r->args.emplace_back("self", nullptr, handle(), /*convert=*/true, /*none=*/false);
+    }
+}
+
+/// Process a keyword argument attribute (*without* a default value)
+template <>
+struct process_attribute<arg> : process_attribute_default<arg> {
+    static void init(const arg &a, function_record *r) {
+        append_self_arg_if_needed(r);
+        r->args.emplace_back(a.name, nullptr, handle(), !a.flag_noconvert, a.flag_none);
+
+        check_kw_only_arg(a, r);
+    }
+};
+
+/// Process a keyword argument attribute (*with* a default value)
+template <>
+struct process_attribute<arg_v> : process_attribute_default<arg_v> {
+    static void init(const arg_v &a, function_record *r) {
+        if (r->is_method && r->args.empty()) {
+            r->args.emplace_back(
+                "self", /*descr=*/nullptr, /*parent=*/handle(), /*convert=*/true, /*none=*/false);
+        }
+
+        if (!a.value) {
+#if defined(PYBIND11_DETAILED_ERROR_MESSAGES)
+            std::string descr("'");
+            if (a.name) {
+                descr += std::string(a.name) + ": ";
+            }
+            descr += a.type + "'";
+            if (r->is_method) {
+                if (r->name) {
+                    descr += " in method '" + (std::string) str(r->scope) + "."
+                             + (std::string) r->name + "'";
+                } else {
+                    descr += " in method of '" + (std::string) str(r->scope) + "'";
+                }
+            } else if (r->name) {
+                descr += " in function '" + (std::string) r->name + "'";
+            }
+            pybind11_fail("arg(): could not convert default argument " + descr
+                          + " into a Python object (type not registered yet?)");
+#else
+            pybind11_fail("arg(): could not convert default argument "
+                          "into a Python object (type not registered yet?). "
+                          "#define PYBIND11_DETAILED_ERROR_MESSAGES or compile in debug mode for "
+                          "more information.");
+#endif
+        }
+        r->args.emplace_back(a.name, a.descr, a.value.inc_ref(), !a.flag_noconvert, a.flag_none);
+
+        check_kw_only_arg(a, r);
+    }
+};
+
+/// Process a keyword-only-arguments-follow pseudo argument
+template <>
+struct process_attribute<kw_only> : process_attribute_default<kw_only> {
+    static void init(const kw_only &, function_record *r) {
+        append_self_arg_if_needed(r);
+        if (r->has_args && r->nargs_pos != static_cast<std::uint16_t>(r->args.size())) {
+            pybind11_fail("Mismatched args() and kw_only(): they must occur at the same relative "
+                          "argument location (or omit kw_only() entirely)");
+        }
+        r->nargs_pos = static_cast<std::uint16_t>(r->args.size());
+    }
+};
+
+/// Process a positional-only-argument maker
+template <>
+struct process_attribute<pos_only> : process_attribute_default<pos_only> {
+    static void init(const pos_only &, function_record *r) {
+        append_self_arg_if_needed(r);
+        r->nargs_pos_only = static_cast<std::uint16_t>(r->args.size());
+        if (r->nargs_pos_only > r->nargs_pos) {
+            pybind11_fail("pos_only(): cannot follow a py::args() argument");
+        }
+        // It also can't follow a kw_only, but a static_assert in pybind11.h checks that
+    }
+};
+
+/// Process a parent class attribute.  Single inheritance only (class_ itself already guarantees
+/// that)
+template <typename T>
+struct process_attribute<T, enable_if_t<is_pyobject<T>::value>>
+    : process_attribute_default<handle> {
+    static void init(const handle &h, type_record *r) { r->bases.append(h); }
+};
+
+/// Process a parent class attribute (deprecated, does not support multiple inheritance)
+template <typename T>
+struct process_attribute<base<T>> : process_attribute_default<base<T>> {
+    static void init(const base<T> &, type_record *r) { r->add_base(typeid(T), nullptr); }
+};
+
+/// Process a multiple inheritance attribute
+template <>
+struct process_attribute<multiple_inheritance> : process_attribute_default<multiple_inheritance> {
+    static void init(const multiple_inheritance &, type_record *r) {
+        r->multiple_inheritance = true;
+    }
+};
+
+template <>
+struct process_attribute<dynamic_attr> : process_attribute_default<dynamic_attr> {
+    static void init(const dynamic_attr &, type_record *r) { r->dynamic_attr = true; }
+};
+
+template <>
+struct process_attribute<custom_type_setup> {
+    static void init(const custom_type_setup &value, type_record *r) {
+        r->custom_type_setup_callback = value.value;
+    }
+};
+
+template <>
+struct process_attribute<is_final> : process_attribute_default<is_final> {
+    static void init(const is_final &, type_record *r) { r->is_final = true; }
+};
+
+template <>
+struct process_attribute<buffer_protocol> : process_attribute_default<buffer_protocol> {
+    static void init(const buffer_protocol &, type_record *r) { r->buffer_protocol = true; }
+};
+
+template <>
+struct process_attribute<metaclass> : process_attribute_default<metaclass> {
+    static void init(const metaclass &m, type_record *r) { r->metaclass = m.value; }
+};
+
+template <>
+struct process_attribute<module_local> : process_attribute_default<module_local> {
+    static void init(const module_local &l, type_record *r) { r->module_local = l.value; }
+};
+
+/// Process a 'prepend' attribute, putting this at the beginning of the overload chain
+template <>
+struct process_attribute<prepend> : process_attribute_default<prepend> {
+    static void init(const prepend &, function_record *r) { r->prepend = true; }
+};
+
+/// Process an 'arithmetic' attribute for enums (does nothing here)
+template <>
+struct process_attribute<arithmetic> : process_attribute_default<arithmetic> {};
+
+template <typename... Ts>
+struct process_attribute<call_guard<Ts...>> : process_attribute_default<call_guard<Ts...>> {};
+
+/**
+ * Process a keep_alive call policy -- invokes keep_alive_impl during the
+ * pre-call handler if both Nurse, Patient != 0 and use the post-call handler
+ * otherwise
+ */
+template <size_t Nurse, size_t Patient>
+struct process_attribute<keep_alive<Nurse, Patient>>
+    : public process_attribute_default<keep_alive<Nurse, Patient>> {
+    template <size_t N = Nurse, size_t P = Patient, enable_if_t<N != 0 && P != 0, int> = 0>
+    static void precall(function_call &call) {
+        keep_alive_impl(Nurse, Patient, call, handle());
+    }
+    template <size_t N = Nurse, size_t P = Patient, enable_if_t<N != 0 && P != 0, int> = 0>
+    static void postcall(function_call &, handle) {}
+    template <size_t N = Nurse, size_t P = Patient, enable_if_t<N == 0 || P == 0, int> = 0>
+    static void precall(function_call &) {}
+    template <size_t N = Nurse, size_t P = Patient, enable_if_t<N == 0 || P == 0, int> = 0>
+    static void postcall(function_call &call, handle ret) {
+        keep_alive_impl(Nurse, Patient, call, ret);
+    }
+};
+
+/// Recursively iterate over variadic template arguments
+template <typename... Args>
+struct process_attributes {
+    static void init(const Args &...args, function_record *r) {
+        PYBIND11_WORKAROUND_INCORRECT_MSVC_C4100(r);
+        PYBIND11_WORKAROUND_INCORRECT_GCC_UNUSED_BUT_SET_PARAMETER(r);
+        using expander = int[];
+        (void) expander{
+            0, ((void) process_attribute<typename std::decay<Args>::type>::init(args, r), 0)...};
+    }
+    static void init(const Args &...args, type_record *r) {
+        PYBIND11_WORKAROUND_INCORRECT_MSVC_C4100(r);
+        PYBIND11_WORKAROUND_INCORRECT_GCC_UNUSED_BUT_SET_PARAMETER(r);
+        using expander = int[];
+        (void) expander{0,
+                        (process_attribute<typename std::decay<Args>::type>::init(args, r), 0)...};
+    }
+    static void precall(function_call &call) {
+        PYBIND11_WORKAROUND_INCORRECT_MSVC_C4100(call);
+        using expander = int[];
+        (void) expander{0,
+                        (process_attribute<typename std::decay<Args>::type>::precall(call), 0)...};
+    }
+    static void postcall(function_call &call, handle fn_ret) {
+        PYBIND11_WORKAROUND_INCORRECT_MSVC_C4100(call, fn_ret);
+        PYBIND11_WORKAROUND_INCORRECT_GCC_UNUSED_BUT_SET_PARAMETER(fn_ret);
+        using expander = int[];
+        (void) expander{
+            0, (process_attribute<typename std::decay<Args>::type>::postcall(call, fn_ret), 0)...};
+    }
+};
+
+template <typename T>
+using is_call_guard = is_instantiation<call_guard, T>;
+
+/// Extract the ``type`` from the first `call_guard` in `Extras...` (or `void_type` if none found)
+template <typename... Extra>
+using extract_guard_t = typename exactly_one_t<is_call_guard, call_guard<>, Extra...>::type;
+
+/// Check the number of named arguments at compile time
+template <typename... Extra,
+          size_t named = constexpr_sum(std::is_base_of<arg, Extra>::value...),
+          size_t self = constexpr_sum(std::is_same<is_method, Extra>::value...)>
+constexpr bool expected_num_args(size_t nargs, bool has_args, bool has_kwargs) {
+    PYBIND11_WORKAROUND_INCORRECT_MSVC_C4100(nargs, has_args, has_kwargs);
+    return named == 0 || (self + named + size_t(has_args) + size_t(has_kwargs)) == nargs;
+}
+
+PYBIND11_NAMESPACE_END(detail)
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/buffer_info.h

@@ -0,0 +1,208 @@
+/*
+    pybind11/buffer_info.h: Python buffer object interface
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "detail/common.h"
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+// Default, C-style strides
+inline std::vector<ssize_t> c_strides(const std::vector<ssize_t> &shape, ssize_t itemsize) {
+    auto ndim = shape.size();
+    std::vector<ssize_t> strides(ndim, itemsize);
+    if (ndim > 0) {
+        for (size_t i = ndim - 1; i > 0; --i) {
+            strides[i - 1] = strides[i] * shape[i];
+        }
+    }
+    return strides;
+}
+
+// F-style strides; default when constructing an array_t with `ExtraFlags & f_style`
+inline std::vector<ssize_t> f_strides(const std::vector<ssize_t> &shape, ssize_t itemsize) {
+    auto ndim = shape.size();
+    std::vector<ssize_t> strides(ndim, itemsize);
+    for (size_t i = 1; i < ndim; ++i) {
+        strides[i] = strides[i - 1] * shape[i - 1];
+    }
+    return strides;
+}
+
+template <typename T, typename SFINAE = void>
+struct compare_buffer_info;
+
+PYBIND11_NAMESPACE_END(detail)
+
+/// Information record describing a Python buffer object
+struct buffer_info {
+    void *ptr = nullptr;          // Pointer to the underlying storage
+    ssize_t itemsize = 0;         // Size of individual items in bytes
+    ssize_t size = 0;             // Total number of entries
+    std::string format;           // For homogeneous buffers, this should be set to
+                                  // format_descriptor<T>::format()
+    ssize_t ndim = 0;             // Number of dimensions
+    std::vector<ssize_t> shape;   // Shape of the tensor (1 entry per dimension)
+    std::vector<ssize_t> strides; // Number of bytes between adjacent entries
+                                  // (for each per dimension)
+    bool readonly = false;        // flag to indicate if the underlying storage may be written to
+
+    buffer_info() = default;
+
+    buffer_info(void *ptr,
+                ssize_t itemsize,
+                const std::string &format,
+                ssize_t ndim,
+                detail::any_container<ssize_t> shape_in,
+                detail::any_container<ssize_t> strides_in,
+                bool readonly = false)
+        : ptr(ptr), itemsize(itemsize), size(1), format(format), ndim(ndim),
+          shape(std::move(shape_in)), strides(std::move(strides_in)), readonly(readonly) {
+        if (ndim != (ssize_t) shape.size() || ndim != (ssize_t) strides.size()) {
+            pybind11_fail("buffer_info: ndim doesn't match shape and/or strides length");
+        }
+        for (size_t i = 0; i < (size_t) ndim; ++i) {
+            size *= shape[i];
+        }
+    }
+
+    template <typename T>
+    buffer_info(T *ptr,
+                detail::any_container<ssize_t> shape_in,
+                detail::any_container<ssize_t> strides_in,
+                bool readonly = false)
+        : buffer_info(private_ctr_tag(),
+                      ptr,
+                      sizeof(T),
+                      format_descriptor<T>::format(),
+                      static_cast<ssize_t>(shape_in->size()),
+                      std::move(shape_in),
+                      std::move(strides_in),
+                      readonly) {}
+
+    buffer_info(void *ptr,
+                ssize_t itemsize,
+                const std::string &format,
+                ssize_t size,
+                bool readonly = false)
+        : buffer_info(ptr, itemsize, format, 1, {size}, {itemsize}, readonly) {}
+
+    template <typename T>
+    buffer_info(T *ptr, ssize_t size, bool readonly = false)
+        : buffer_info(ptr, sizeof(T), format_descriptor<T>::format(), size, readonly) {}
+
+    template <typename T>
+    buffer_info(const T *ptr, ssize_t size, bool readonly = true)
+        : buffer_info(
+              const_cast<T *>(ptr), sizeof(T), format_descriptor<T>::format(), size, readonly) {}
+
+    explicit buffer_info(Py_buffer *view, bool ownview = true)
+        : buffer_info(
+              view->buf,
+              view->itemsize,
+              view->format,
+              view->ndim,
+              {view->shape, view->shape + view->ndim},
+              /* Though buffer::request() requests PyBUF_STRIDES, ctypes objects
+               * ignore this flag and return a view with NULL strides.
+               * When strides are NULL, build them manually.  */
+              view->strides
+                  ? std::vector<ssize_t>(view->strides, view->strides + view->ndim)
+                  : detail::c_strides({view->shape, view->shape + view->ndim}, view->itemsize),
+              (view->readonly != 0)) {
+        // NOLINTNEXTLINE(cppcoreguidelines-prefer-member-initializer)
+        this->m_view = view;
+        // NOLINTNEXTLINE(cppcoreguidelines-prefer-member-initializer)
+        this->ownview = ownview;
+    }
+
+    buffer_info(const buffer_info &) = delete;
+    buffer_info &operator=(const buffer_info &) = delete;
+
+    buffer_info(buffer_info &&other) noexcept { (*this) = std::move(other); }
+
+    buffer_info &operator=(buffer_info &&rhs) noexcept {
+        ptr = rhs.ptr;
+        itemsize = rhs.itemsize;
+        size = rhs.size;
+        format = std::move(rhs.format);
+        ndim = rhs.ndim;
+        shape = std::move(rhs.shape);
+        strides = std::move(rhs.strides);
+        std::swap(m_view, rhs.m_view);
+        std::swap(ownview, rhs.ownview);
+        readonly = rhs.readonly;
+        return *this;
+    }
+
+    ~buffer_info() {
+        if (m_view && ownview) {
+            PyBuffer_Release(m_view);
+            delete m_view;
+        }
+    }
+
+    Py_buffer *view() const { return m_view; }
+    Py_buffer *&view() { return m_view; }
+
+    /* True if the buffer item type is equivalent to `T`. */
+    // To define "equivalent" by example:
+    // `buffer_info::item_type_is_equivalent_to<int>(b)` and
+    // `buffer_info::item_type_is_equivalent_to<long>(b)` may both be true
+    // on some platforms, but `int` and `unsigned` will never be equivalent.
+    // For the ground truth, please inspect `detail::compare_buffer_info<>`.
+    template <typename T>
+    bool item_type_is_equivalent_to() const {
+        return detail::compare_buffer_info<T>::compare(*this);
+    }
+
+private:
+    struct private_ctr_tag {};
+
+    buffer_info(private_ctr_tag,
+                void *ptr,
+                ssize_t itemsize,
+                const std::string &format,
+                ssize_t ndim,
+                detail::any_container<ssize_t> &&shape_in,
+                detail::any_container<ssize_t> &&strides_in,
+                bool readonly)
+        : buffer_info(
+              ptr, itemsize, format, ndim, std::move(shape_in), std::move(strides_in), readonly) {}
+
+    Py_buffer *m_view = nullptr;
+    bool ownview = false;
+};
+
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+template <typename T, typename SFINAE>
+struct compare_buffer_info {
+    static bool compare(const buffer_info &b) {
+        // NOLINTNEXTLINE(bugprone-sizeof-expression) Needed for `PyObject *`
+        return b.format == format_descriptor<T>::format() && b.itemsize == (ssize_t) sizeof(T);
+    }
+};
+
+template <typename T>
+struct compare_buffer_info<T, detail::enable_if_t<std::is_integral<T>::value>> {
+    static bool compare(const buffer_info &b) {
+        return (size_t) b.itemsize == sizeof(T)
+               && (b.format == format_descriptor<T>::value
+                   || ((sizeof(T) == sizeof(long))
+                       && b.format == (std::is_unsigned<T>::value ? "L" : "l"))
+                   || ((sizeof(T) == sizeof(size_t))
+                       && b.format == (std::is_unsigned<T>::value ? "N" : "n")));
+    }
+};
+
+PYBIND11_NAMESPACE_END(detail)
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/cast.h

@@ -0,0 +1,1855 @@
+/*
+    pybind11/cast.h: Partial template specializations to cast between
+    C++ and Python types
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "detail/common.h"
+#include "detail/descr.h"
+#include "detail/type_caster_base.h"
+#include "detail/typeid.h"
+#include "pytypes.h"
+
+#include <array>
+#include <cstring>
+#include <functional>
+#include <iosfwd>
+#include <iterator>
+#include <memory>
+#include <string>
+#include <tuple>
+#include <type_traits>
+#include <utility>
+#include <vector>
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+PYBIND11_WARNING_DISABLE_MSVC(4127)
+
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+template <typename type, typename SFINAE = void>
+class type_caster : public type_caster_base<type> {};
+template <typename type>
+using make_caster = type_caster<intrinsic_t<type>>;
+
+// Shortcut for calling a caster's `cast_op_type` cast operator for casting a type_caster to a T
+template <typename T>
+typename make_caster<T>::template cast_op_type<T> cast_op(make_caster<T> &caster) {
+    using result_t = typename make_caster<T>::template cast_op_type<T>; // See PR #4893
+    return caster.operator result_t();
+}
+template <typename T>
+typename make_caster<T>::template cast_op_type<typename std::add_rvalue_reference<T>::type>
+cast_op(make_caster<T> &&caster) {
+    using result_t = typename make_caster<T>::template cast_op_type<
+        typename std::add_rvalue_reference<T>::type>; // See PR #4893
+    return std::move(caster).operator result_t();
+}
+
+template <typename type>
+class type_caster<std::reference_wrapper<type>> {
+private:
+    using caster_t = make_caster<type>;
+    caster_t subcaster;
+    using reference_t = type &;
+    using subcaster_cast_op_type = typename caster_t::template cast_op_type<reference_t>;
+
+    static_assert(
+        std::is_same<typename std::remove_const<type>::type &, subcaster_cast_op_type>::value
+            || std::is_same<reference_t, subcaster_cast_op_type>::value,
+        "std::reference_wrapper<T> caster requires T to have a caster with an "
+        "`operator T &()` or `operator const T &()`");
+
+public:
+    bool load(handle src, bool convert) { return subcaster.load(src, convert); }
+    static constexpr auto name = caster_t::name;
+    static handle
+    cast(const std::reference_wrapper<type> &src, return_value_policy policy, handle parent) {
+        // It is definitely wrong to take ownership of this pointer, so mask that rvp
+        if (policy == return_value_policy::take_ownership
+            || policy == return_value_policy::automatic) {
+            policy = return_value_policy::automatic_reference;
+        }
+        return caster_t::cast(&src.get(), policy, parent);
+    }
+    template <typename T>
+    using cast_op_type = std::reference_wrapper<type>;
+    explicit operator std::reference_wrapper<type>() { return cast_op<type &>(subcaster); }
+};
+
+#define PYBIND11_TYPE_CASTER(type, py_name)                                                       \
+protected:                                                                                        \
+    type value;                                                                                   \
+                                                                                                  \
+public:                                                                                           \
+    static constexpr auto name = py_name;                                                         \
+    template <typename T_,                                                                        \
+              ::pybind11::detail::enable_if_t<                                                    \
+                  std::is_same<type, ::pybind11::detail::remove_cv_t<T_>>::value,                 \
+                  int>                                                                            \
+              = 0>                                                                                \
+    static ::pybind11::handle cast(                                                               \
+        T_ *src, ::pybind11::return_value_policy policy, ::pybind11::handle parent) {             \
+        if (!src)                                                                                 \
+            return ::pybind11::none().release();                                                  \
+        if (policy == ::pybind11::return_value_policy::take_ownership) {                          \
+            auto h = cast(std::move(*src), policy, parent);                                       \
+            delete src;                                                                           \
+            return h;                                                                             \
+        }                                                                                         \
+        return cast(*src, policy, parent);                                                        \
+    }                                                                                             \
+    operator type *() { return &value; }               /* NOLINT(bugprone-macro-parentheses) */   \
+    operator type &() { return value; }                /* NOLINT(bugprone-macro-parentheses) */   \
+    operator type &&() && { return std::move(value); } /* NOLINT(bugprone-macro-parentheses) */   \
+    template <typename T_>                                                                        \
+    using cast_op_type = ::pybind11::detail::movable_cast_op_type<T_>
+
+template <typename CharT>
+using is_std_char_type = any_of<std::is_same<CharT, char>, /* std::string */
+#if defined(PYBIND11_HAS_U8STRING)
+                                std::is_same<CharT, char8_t>, /* std::u8string */
+#endif
+                                std::is_same<CharT, char16_t>, /* std::u16string */
+                                std::is_same<CharT, char32_t>, /* std::u32string */
+                                std::is_same<CharT, wchar_t>   /* std::wstring */
+                                >;
+
+template <typename T>
+struct type_caster<T, enable_if_t<std::is_arithmetic<T>::value && !is_std_char_type<T>::value>> {
+    using _py_type_0 = conditional_t<sizeof(T) <= sizeof(long), long, long long>;
+    using _py_type_1 = conditional_t<std::is_signed<T>::value,
+                                     _py_type_0,
+                                     typename std::make_unsigned<_py_type_0>::type>;
+    using py_type = conditional_t<std::is_floating_point<T>::value, double, _py_type_1>;
+
+public:
+    bool load(handle src, bool convert) {
+        py_type py_value;
+
+        if (!src) {
+            return false;
+        }
+
+#if !defined(PYPY_VERSION)
+        auto index_check = [](PyObject *o) { return PyIndex_Check(o); };
+#else
+        // In PyPy 7.3.3, `PyIndex_Check` is implemented by calling `__index__`,
+        // while CPython only considers the existence of `nb_index`/`__index__`.
+        auto index_check = [](PyObject *o) { return hasattr(o, "__index__"); };
+#endif
+
+        if (std::is_floating_point<T>::value) {
+            if (convert || PyFloat_Check(src.ptr())) {
+                py_value = (py_type) PyFloat_AsDouble(src.ptr());
+            } else {
+                return false;
+            }
+        } else if (PyFloat_Check(src.ptr())
+                   || (!convert && !PYBIND11_LONG_CHECK(src.ptr()) && !index_check(src.ptr()))) {
+            return false;
+        } else {
+            handle src_or_index = src;
+            // PyPy: 7.3.7's 3.8 does not implement PyLong_*'s __index__ calls.
+#if PY_VERSION_HEX < 0x03080000 || defined(PYPY_VERSION)
+            object index;
+            if (!PYBIND11_LONG_CHECK(src.ptr())) { // So: index_check(src.ptr())
+                index = reinterpret_steal<object>(PyNumber_Index(src.ptr()));
+                if (!index) {
+                    PyErr_Clear();
+                    if (!convert)
+                        return false;
+                } else {
+                    src_or_index = index;
+                }
+            }
+#endif
+            if (std::is_unsigned<py_type>::value) {
+                py_value = as_unsigned<py_type>(src_or_index.ptr());
+            } else { // signed integer:
+                py_value = sizeof(T) <= sizeof(long)
+                               ? (py_type) PyLong_AsLong(src_or_index.ptr())
+                               : (py_type) PYBIND11_LONG_AS_LONGLONG(src_or_index.ptr());
+            }
+        }
+
+        // Python API reported an error
+        bool py_err = py_value == (py_type) -1 && PyErr_Occurred();
+
+        // Check to see if the conversion is valid (integers should match exactly)
+        // Signed/unsigned checks happen elsewhere
+        if (py_err
+            || (std::is_integral<T>::value && sizeof(py_type) != sizeof(T)
+                && py_value != (py_type) (T) py_value)) {
+            PyErr_Clear();
+            if (py_err && convert && (PyNumber_Check(src.ptr()) != 0)) {
+                auto tmp = reinterpret_steal<object>(std::is_floating_point<T>::value
+                                                         ? PyNumber_Float(src.ptr())
+                                                         : PyNumber_Long(src.ptr()));
+                PyErr_Clear();
+                return load(tmp, false);
+            }
+            return false;
+        }
+
+        value = (T) py_value;
+        return true;
+    }
+
+    template <typename U = T>
+    static typename std::enable_if<std::is_floating_point<U>::value, handle>::type
+    cast(U src, return_value_policy /* policy */, handle /* parent */) {
+        return PyFloat_FromDouble((double) src);
+    }
+
+    template <typename U = T>
+    static typename std::enable_if<!std::is_floating_point<U>::value && std::is_signed<U>::value
+                                       && (sizeof(U) <= sizeof(long)),
+                                   handle>::type
+    cast(U src, return_value_policy /* policy */, handle /* parent */) {
+        return PYBIND11_LONG_FROM_SIGNED((long) src);
+    }
+
+    template <typename U = T>
+    static typename std::enable_if<!std::is_floating_point<U>::value && std::is_unsigned<U>::value
+                                       && (sizeof(U) <= sizeof(unsigned long)),
+                                   handle>::type
+    cast(U src, return_value_policy /* policy */, handle /* parent */) {
+        return PYBIND11_LONG_FROM_UNSIGNED((unsigned long) src);
+    }
+
+    template <typename U = T>
+    static typename std::enable_if<!std::is_floating_point<U>::value && std::is_signed<U>::value
+                                       && (sizeof(U) > sizeof(long)),
+                                   handle>::type
+    cast(U src, return_value_policy /* policy */, handle /* parent */) {
+        return PyLong_FromLongLong((long long) src);
+    }
+
+    template <typename U = T>
+    static typename std::enable_if<!std::is_floating_point<U>::value && std::is_unsigned<U>::value
+                                       && (sizeof(U) > sizeof(unsigned long)),
+                                   handle>::type
+    cast(U src, return_value_policy /* policy */, handle /* parent */) {
+        return PyLong_FromUnsignedLongLong((unsigned long long) src);
+    }
+
+    PYBIND11_TYPE_CASTER(T, const_name<std::is_integral<T>::value>("int", "float"));
+};
+
+template <typename T>
+struct void_caster {
+public:
+    bool load(handle src, bool) {
+        if (src && src.is_none()) {
+            return true;
+        }
+        return false;
+    }
+    static handle cast(T, return_value_policy /* policy */, handle /* parent */) {
+        return none().release();
+    }
+    PYBIND11_TYPE_CASTER(T, const_name("None"));
+};
+
+template <>
+class type_caster<void_type> : public void_caster<void_type> {};
+
+template <>
+class type_caster<void> : public type_caster<void_type> {
+public:
+    using type_caster<void_type>::cast;
+
+    bool load(handle h, bool) {
+        if (!h) {
+            return false;
+        }
+        if (h.is_none()) {
+            value = nullptr;
+            return true;
+        }
+
+        /* Check if this is a capsule */
+        if (isinstance<capsule>(h)) {
+            value = reinterpret_borrow<capsule>(h);
+            return true;
+        }
+
+        /* Check if this is a C++ type */
+        const auto &bases = all_type_info((PyTypeObject *) type::handle_of(h).ptr());
+        if (bases.size() == 1) { // Only allowing loading from a single-value type
+            value = values_and_holders(reinterpret_cast<instance *>(h.ptr())).begin()->value_ptr();
+            return true;
+        }
+
+        /* Fail */
+        return false;
+    }
+
+    static handle cast(const void *ptr, return_value_policy /* policy */, handle /* parent */) {
+        if (ptr) {
+            return capsule(ptr).release();
+        }
+        return none().release();
+    }
+
+    template <typename T>
+    using cast_op_type = void *&;
+    explicit operator void *&() { return value; }
+    static constexpr auto name = const_name("capsule");
+
+private:
+    void *value = nullptr;
+};
+
+template <>
+class type_caster<std::nullptr_t> : public void_caster<std::nullptr_t> {};
+
+template <>
+class type_caster<bool> {
+public:
+    bool load(handle src, bool convert) {
+        if (!src) {
+            return false;
+        }
+        if (src.ptr() == Py_True) {
+            value = true;
+            return true;
+        }
+        if (src.ptr() == Py_False) {
+            value = false;
+            return true;
+        }
+        if (convert || is_numpy_bool(src)) {
+            // (allow non-implicit conversion for numpy booleans), use strncmp
+            // since NumPy 1.x had an additional trailing underscore.
+
+            Py_ssize_t res = -1;
+            if (src.is_none()) {
+                res = 0; // None is implicitly converted to False
+            }
+#if defined(PYPY_VERSION)
+            // On PyPy, check that "__bool__" attr exists
+            else if (hasattr(src, PYBIND11_BOOL_ATTR)) {
+                res = PyObject_IsTrue(src.ptr());
+            }
+#else
+            // Alternate approach for CPython: this does the same as the above, but optimized
+            // using the CPython API so as to avoid an unneeded attribute lookup.
+            else if (auto *tp_as_number = src.ptr()->ob_type->tp_as_number) {
+                if (PYBIND11_NB_BOOL(tp_as_number)) {
+                    res = (*PYBIND11_NB_BOOL(tp_as_number))(src.ptr());
+                }
+            }
+#endif
+            if (res == 0 || res == 1) {
+                value = (res != 0);
+                return true;
+            }
+            PyErr_Clear();
+        }
+        return false;
+    }
+    static handle cast(bool src, return_value_policy /* policy */, handle /* parent */) {
+        return handle(src ? Py_True : Py_False).inc_ref();
+    }
+    PYBIND11_TYPE_CASTER(bool, const_name("bool"));
+
+private:
+    // Test if an object is a NumPy boolean (without fetching the type).
+    static inline bool is_numpy_bool(handle object) {
+        const char *type_name = Py_TYPE(object.ptr())->tp_name;
+        // Name changed to `numpy.bool` in NumPy 2, `numpy.bool_` is needed for 1.x support
+        return std::strcmp("numpy.bool", type_name) == 0
+               || std::strcmp("numpy.bool_", type_name) == 0;
+    }
+};
+
+// Helper class for UTF-{8,16,32} C++ stl strings:
+template <typename StringType, bool IsView = false>
+struct string_caster {
+    using CharT = typename StringType::value_type;
+
+    // Simplify life by being able to assume standard char sizes (the standard only guarantees
+    // minimums, but Python requires exact sizes)
+    static_assert(!std::is_same<CharT, char>::value || sizeof(CharT) == 1,
+                  "Unsupported char size != 1");
+#if defined(PYBIND11_HAS_U8STRING)
+    static_assert(!std::is_same<CharT, char8_t>::value || sizeof(CharT) == 1,
+                  "Unsupported char8_t size != 1");
+#endif
+    static_assert(!std::is_same<CharT, char16_t>::value || sizeof(CharT) == 2,
+                  "Unsupported char16_t size != 2");
+    static_assert(!std::is_same<CharT, char32_t>::value || sizeof(CharT) == 4,
+                  "Unsupported char32_t size != 4");
+    // wchar_t can be either 16 bits (Windows) or 32 (everywhere else)
+    static_assert(!std::is_same<CharT, wchar_t>::value || sizeof(CharT) == 2 || sizeof(CharT) == 4,
+                  "Unsupported wchar_t size != 2/4");
+    static constexpr size_t UTF_N = 8 * sizeof(CharT);
+
+    bool load(handle src, bool) {
+        handle load_src = src;
+        if (!src) {
+            return false;
+        }
+        if (!PyUnicode_Check(load_src.ptr())) {
+            return load_raw(load_src);
+        }
+
+        // For UTF-8 we avoid the need for a temporary `bytes` object by using
+        // `PyUnicode_AsUTF8AndSize`.
+        if (UTF_N == 8) {
+            Py_ssize_t size = -1;
+            const auto *buffer
+                = reinterpret_cast<const CharT *>(PyUnicode_AsUTF8AndSize(load_src.ptr(), &size));
+            if (!buffer) {
+                PyErr_Clear();
+                return false;
+            }
+            value = StringType(buffer, static_cast<size_t>(size));
+            return true;
+        }
+
+        auto utfNbytes
+            = reinterpret_steal<object>(PyUnicode_AsEncodedString(load_src.ptr(),
+                                                                  UTF_N == 8    ? "utf-8"
+                                                                  : UTF_N == 16 ? "utf-16"
+                                                                                : "utf-32",
+                                                                  nullptr));
+        if (!utfNbytes) {
+            PyErr_Clear();
+            return false;
+        }
+
+        const auto *buffer
+            = reinterpret_cast<const CharT *>(PYBIND11_BYTES_AS_STRING(utfNbytes.ptr()));
+        size_t length = (size_t) PYBIND11_BYTES_SIZE(utfNbytes.ptr()) / sizeof(CharT);
+        // Skip BOM for UTF-16/32
+        if (UTF_N > 8) {
+            buffer++;
+            length--;
+        }
+        value = StringType(buffer, length);
+
+        // If we're loading a string_view we need to keep the encoded Python object alive:
+        if (IsView) {
+            loader_life_support::add_patient(utfNbytes);
+        }
+
+        return true;
+    }
+
+    static handle
+    cast(const StringType &src, return_value_policy /* policy */, handle /* parent */) {
+        const char *buffer = reinterpret_cast<const char *>(src.data());
+        auto nbytes = ssize_t(src.size() * sizeof(CharT));
+        handle s = decode_utfN(buffer, nbytes);
+        if (!s) {
+            throw error_already_set();
+        }
+        return s;
+    }
+
+    PYBIND11_TYPE_CASTER(StringType, const_name(PYBIND11_STRING_NAME));
+
+private:
+    static handle decode_utfN(const char *buffer, ssize_t nbytes) {
+#if !defined(PYPY_VERSION)
+        return UTF_N == 8    ? PyUnicode_DecodeUTF8(buffer, nbytes, nullptr)
+               : UTF_N == 16 ? PyUnicode_DecodeUTF16(buffer, nbytes, nullptr, nullptr)
+                             : PyUnicode_DecodeUTF32(buffer, nbytes, nullptr, nullptr);
+#else
+        // PyPy segfaults when on PyUnicode_DecodeUTF16 (and possibly on PyUnicode_DecodeUTF32 as
+        // well), so bypass the whole thing by just passing the encoding as a string value, which
+        // works properly:
+        return PyUnicode_Decode(buffer,
+                                nbytes,
+                                UTF_N == 8    ? "utf-8"
+                                : UTF_N == 16 ? "utf-16"
+                                              : "utf-32",
+                                nullptr);
+#endif
+    }
+
+    // When loading into a std::string or char*, accept a bytes/bytearray object as-is (i.e.
+    // without any encoding/decoding attempt).  For other C++ char sizes this is a no-op.
+    // which supports loading a unicode from a str, doesn't take this path.
+    template <typename C = CharT>
+    bool load_raw(enable_if_t<std::is_same<C, char>::value, handle> src) {
+        if (PYBIND11_BYTES_CHECK(src.ptr())) {
+            // We were passed raw bytes; accept it into a std::string or char*
+            // without any encoding attempt.
+            const char *bytes = PYBIND11_BYTES_AS_STRING(src.ptr());
+            if (!bytes) {
+                pybind11_fail("Unexpected PYBIND11_BYTES_AS_STRING() failure.");
+            }
+            value = StringType(bytes, (size_t) PYBIND11_BYTES_SIZE(src.ptr()));
+            return true;
+        }
+        if (PyByteArray_Check(src.ptr())) {
+            // We were passed a bytearray; accept it into a std::string or char*
+            // without any encoding attempt.
+            const char *bytearray = PyByteArray_AsString(src.ptr());
+            if (!bytearray) {
+                pybind11_fail("Unexpected PyByteArray_AsString() failure.");
+            }
+            value = StringType(bytearray, (size_t) PyByteArray_Size(src.ptr()));
+            return true;
+        }
+
+        return false;
+    }
+
+    template <typename C = CharT>
+    bool load_raw(enable_if_t<!std::is_same<C, char>::value, handle>) {
+        return false;
+    }
+};
+
+template <typename CharT, class Traits, class Allocator>
+struct type_caster<std::basic_string<CharT, Traits, Allocator>,
+                   enable_if_t<is_std_char_type<CharT>::value>>
+    : string_caster<std::basic_string<CharT, Traits, Allocator>> {};
+
+#ifdef PYBIND11_HAS_STRING_VIEW
+template <typename CharT, class Traits>
+struct type_caster<std::basic_string_view<CharT, Traits>,
+                   enable_if_t<is_std_char_type<CharT>::value>>
+    : string_caster<std::basic_string_view<CharT, Traits>, true> {};
+#endif
+
+// Type caster for C-style strings.  We basically use a std::string type caster, but also add the
+// ability to use None as a nullptr char* (which the string caster doesn't allow).
+template <typename CharT>
+struct type_caster<CharT, enable_if_t<is_std_char_type<CharT>::value>> {
+    using StringType = std::basic_string<CharT>;
+    using StringCaster = make_caster<StringType>;
+    StringCaster str_caster;
+    bool none = false;
+    CharT one_char = 0;
+
+public:
+    bool load(handle src, bool convert) {
+        if (!src) {
+            return false;
+        }
+        if (src.is_none()) {
+            // Defer accepting None to other overloads (if we aren't in convert mode):
+            if (!convert) {
+                return false;
+            }
+            none = true;
+            return true;
+        }
+        return str_caster.load(src, convert);
+    }
+
+    static handle cast(const CharT *src, return_value_policy policy, handle parent) {
+        if (src == nullptr) {
+            return pybind11::none().release();
+        }
+        return StringCaster::cast(StringType(src), policy, parent);
+    }
+
+    static handle cast(CharT src, return_value_policy policy, handle parent) {
+        if (std::is_same<char, CharT>::value) {
+            handle s = PyUnicode_DecodeLatin1((const char *) &src, 1, nullptr);
+            if (!s) {
+                throw error_already_set();
+            }
+            return s;
+        }
+        return StringCaster::cast(StringType(1, src), policy, parent);
+    }
+
+    explicit operator CharT *() {
+        return none ? nullptr : const_cast<CharT *>(static_cast<StringType &>(str_caster).c_str());
+    }
+    explicit operator CharT &() {
+        if (none) {
+            throw value_error("Cannot convert None to a character");
+        }
+
+        auto &value = static_cast<StringType &>(str_caster);
+        size_t str_len = value.size();
+        if (str_len == 0) {
+            throw value_error("Cannot convert empty string to a character");
+        }
+
+        // If we're in UTF-8 mode, we have two possible failures: one for a unicode character that
+        // is too high, and one for multiple unicode characters (caught later), so we need to
+        // figure out how long the first encoded character is in bytes to distinguish between these
+        // two errors.  We also allow want to allow unicode characters U+0080 through U+00FF, as
+        // those can fit into a single char value.
+        if (StringCaster::UTF_N == 8 && str_len > 1 && str_len <= 4) {
+            auto v0 = static_cast<unsigned char>(value[0]);
+            // low bits only: 0-127
+            // 0b110xxxxx - start of 2-byte sequence
+            // 0b1110xxxx - start of 3-byte sequence
+            // 0b11110xxx - start of 4-byte sequence
+            size_t char0_bytes = (v0 & 0x80) == 0      ? 1
+                                 : (v0 & 0xE0) == 0xC0 ? 2
+                                 : (v0 & 0xF0) == 0xE0 ? 3
+                                                       : 4;
+
+            if (char0_bytes == str_len) {
+                // If we have a 128-255 value, we can decode it into a single char:
+                if (char0_bytes == 2 && (v0 & 0xFC) == 0xC0) { // 0x110000xx 0x10xxxxxx
+                    one_char = static_cast<CharT>(((v0 & 3) << 6)
+                                                  + (static_cast<unsigned char>(value[1]) & 0x3F));
+                    return one_char;
+                }
+                // Otherwise we have a single character, but it's > U+00FF
+                throw value_error("Character code point not in range(0x100)");
+            }
+        }
+
+        // UTF-16 is much easier: we can only have a surrogate pair for values above U+FFFF, thus a
+        // surrogate pair with total length 2 instantly indicates a range error (but not a "your
+        // string was too long" error).
+        else if (StringCaster::UTF_N == 16 && str_len == 2) {
+            one_char = static_cast<CharT>(value[0]);
+            if (one_char >= 0xD800 && one_char < 0xE000) {
+                throw value_error("Character code point not in range(0x10000)");
+            }
+        }
+
+        if (str_len != 1) {
+            throw value_error("Expected a character, but multi-character string found");
+        }
+
+        one_char = value[0];
+        return one_char;
+    }
+
+    static constexpr auto name = const_name(PYBIND11_STRING_NAME);
+    template <typename _T>
+    using cast_op_type = pybind11::detail::cast_op_type<_T>;
+};
+
+// Base implementation for std::tuple and std::pair
+template <template <typename...> class Tuple, typename... Ts>
+class tuple_caster {
+    using type = Tuple<Ts...>;
+    static constexpr auto size = sizeof...(Ts);
+    using indices = make_index_sequence<size>;
+
+public:
+    bool load(handle src, bool convert) {
+        if (!isinstance<sequence>(src)) {
+            return false;
+        }
+        const auto seq = reinterpret_borrow<sequence>(src);
+        if (seq.size() != size) {
+            return false;
+        }
+        return load_impl(seq, convert, indices{});
+    }
+
+    template <typename T>
+    static handle cast(T &&src, return_value_policy policy, handle parent) {
+        return cast_impl(std::forward<T>(src), policy, parent, indices{});
+    }
+
+    // copied from the PYBIND11_TYPE_CASTER macro
+    template <typename T>
+    static handle cast(T *src, return_value_policy policy, handle parent) {
+        if (!src) {
+            return none().release();
+        }
+        if (policy == return_value_policy::take_ownership) {
+            auto h = cast(std::move(*src), policy, parent);
+            delete src;
+            return h;
+        }
+        return cast(*src, policy, parent);
+    }
+
+    static constexpr auto name = const_name("tuple[")
+                                 + ::pybind11::detail::concat(make_caster<Ts>::name...)
+                                 + const_name("]");
+
+    template <typename T>
+    using cast_op_type = type;
+
+    explicit operator type() & { return implicit_cast(indices{}); }
+    explicit operator type() && { return std::move(*this).implicit_cast(indices{}); }
+
+protected:
+    template <size_t... Is>
+    type implicit_cast(index_sequence<Is...>) & {
+        return type(cast_op<Ts>(std::get<Is>(subcasters))...);
+    }
+    template <size_t... Is>
+    type implicit_cast(index_sequence<Is...>) && {
+        return type(cast_op<Ts>(std::move(std::get<Is>(subcasters)))...);
+    }
+
+    static constexpr bool load_impl(const sequence &, bool, index_sequence<>) { return true; }
+
+    template <size_t... Is>
+    bool load_impl(const sequence &seq, bool convert, index_sequence<Is...>) {
+#ifdef __cpp_fold_expressions
+        if ((... || !std::get<Is>(subcasters).load(seq[Is], convert))) {
+            return false;
+        }
+#else
+        for (bool r : {std::get<Is>(subcasters).load(seq[Is], convert)...}) {
+            if (!r) {
+                return false;
+            }
+        }
+#endif
+        return true;
+    }
+
+    /* Implementation: Convert a C++ tuple into a Python tuple */
+    template <typename T, size_t... Is>
+    static handle
+    cast_impl(T &&src, return_value_policy policy, handle parent, index_sequence<Is...>) {
+        PYBIND11_WORKAROUND_INCORRECT_MSVC_C4100(src, policy, parent);
+        PYBIND11_WORKAROUND_INCORRECT_GCC_UNUSED_BUT_SET_PARAMETER(policy, parent);
+        std::array<object, size> entries{{reinterpret_steal<object>(
+            make_caster<Ts>::cast(std::get<Is>(std::forward<T>(src)), policy, parent))...}};
+        for (const auto &entry : entries) {
+            if (!entry) {
+                return handle();
+            }
+        }
+        tuple result(size);
+        int counter = 0;
+        for (auto &entry : entries) {
+            PyTuple_SET_ITEM(result.ptr(), counter++, entry.release().ptr());
+        }
+        return result.release();
+    }
+
+    Tuple<make_caster<Ts>...> subcasters;
+};
+
+template <typename T1, typename T2>
+class type_caster<std::pair<T1, T2>> : public tuple_caster<std::pair, T1, T2> {};
+
+template <typename... Ts>
+class type_caster<std::tuple<Ts...>> : public tuple_caster<std::tuple, Ts...> {};
+
+template <>
+class type_caster<std::tuple<>> : public tuple_caster<std::tuple> {
+public:
+    // PEP 484 specifies this syntax for an empty tuple
+    static constexpr auto name = const_name("tuple[()]");
+};
+
+/// Helper class which abstracts away certain actions. Users can provide specializations for
+/// custom holders, but it's only necessary if the type has a non-standard interface.
+template <typename T>
+struct holder_helper {
+    static auto get(const T &p) -> decltype(p.get()) { return p.get(); }
+};
+
+/// Type caster for holder types like std::shared_ptr, etc.
+/// The SFINAE hook is provided to help work around the current lack of support
+/// for smart-pointer interoperability. Please consider it an implementation
+/// detail that may change in the future, as formal support for smart-pointer
+/// interoperability is added into pybind11.
+template <typename type, typename holder_type, typename SFINAE = void>
+struct copyable_holder_caster : public type_caster_base<type> {
+public:
+    using base = type_caster_base<type>;
+    static_assert(std::is_base_of<base, type_caster<type>>::value,
+                  "Holder classes are only supported for custom types");
+    using base::base;
+    using base::cast;
+    using base::typeinfo;
+    using base::value;
+
+    bool load(handle src, bool convert) {
+        return base::template load_impl<copyable_holder_caster<type, holder_type>>(src, convert);
+    }
+
+    explicit operator type *() { return this->value; }
+    // static_cast works around compiler error with MSVC 17 and CUDA 10.2
+    // see issue #2180
+    explicit operator type &() { return *(static_cast<type *>(this->value)); }
+    explicit operator holder_type *() { return std::addressof(holder); }
+    explicit operator holder_type &() { return holder; }
+
+    static handle cast(const holder_type &src, return_value_policy, handle) {
+        const auto *ptr = holder_helper<holder_type>::get(src);
+        return type_caster_base<type>::cast_holder(ptr, &src);
+    }
+
+protected:
+    friend class type_caster_generic;
+    void check_holder_compat() {
+        if (typeinfo->default_holder) {
+            throw cast_error("Unable to load a custom holder type from a default-holder instance");
+        }
+    }
+
+    void load_value(value_and_holder &&v_h) {
+        if (v_h.holder_constructed()) {
+            value = v_h.value_ptr();
+            holder = v_h.template holder<holder_type>();
+            return;
+        }
+        throw cast_error("Unable to cast from non-held to held instance (T& to Holder<T>) "
+#if !defined(PYBIND11_DETAILED_ERROR_MESSAGES)
+                         "(#define PYBIND11_DETAILED_ERROR_MESSAGES or compile in debug mode for "
+                         "type information)");
+#else
+                         "of type '"
+                         + type_id<holder_type>() + "''");
+#endif
+    }
+
+    template <typename T = holder_type,
+              detail::enable_if_t<!std::is_constructible<T, const T &, type *>::value, int> = 0>
+    bool try_implicit_casts(handle, bool) {
+        return false;
+    }
+
+    template <typename T = holder_type,
+              detail::enable_if_t<std::is_constructible<T, const T &, type *>::value, int> = 0>
+    bool try_implicit_casts(handle src, bool convert) {
+        for (auto &cast : typeinfo->implicit_casts) {
+            copyable_holder_caster sub_caster(*cast.first);
+            if (sub_caster.load(src, convert)) {
+                value = cast.second(sub_caster.value);
+                holder = holder_type(sub_caster.holder, (type *) value);
+                return true;
+            }
+        }
+        return false;
+    }
+
+    static bool try_direct_conversions(handle) { return false; }
+
+    holder_type holder;
+};
+
+/// Specialize for the common std::shared_ptr, so users don't need to
+template <typename T>
+class type_caster<std::shared_ptr<T>> : public copyable_holder_caster<T, std::shared_ptr<T>> {};
+
+/// Type caster for holder types like std::unique_ptr.
+/// Please consider the SFINAE hook an implementation detail, as explained
+/// in the comment for the copyable_holder_caster.
+template <typename type, typename holder_type, typename SFINAE = void>
+struct move_only_holder_caster {
+    static_assert(std::is_base_of<type_caster_base<type>, type_caster<type>>::value,
+                  "Holder classes are only supported for custom types");
+
+    static handle cast(holder_type &&src, return_value_policy, handle) {
+        auto *ptr = holder_helper<holder_type>::get(src);
+        return type_caster_base<type>::cast_holder(ptr, std::addressof(src));
+    }
+    static constexpr auto name = type_caster_base<type>::name;
+};
+
+template <typename type, typename deleter>
+class type_caster<std::unique_ptr<type, deleter>>
+    : public move_only_holder_caster<type, std::unique_ptr<type, deleter>> {};
+
+template <typename type, typename holder_type>
+using type_caster_holder = conditional_t<is_copy_constructible<holder_type>::value,
+                                         copyable_holder_caster<type, holder_type>,
+                                         move_only_holder_caster<type, holder_type>>;
+
+template <typename T, bool Value = false>
+struct always_construct_holder {
+    static constexpr bool value = Value;
+};
+
+/// Create a specialization for custom holder types (silently ignores std::shared_ptr)
+#define PYBIND11_DECLARE_HOLDER_TYPE(type, holder_type, ...)                                      \
+    PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)                                                  \
+    namespace detail {                                                                            \
+    template <typename type>                                                                      \
+    struct always_construct_holder<holder_type> : always_construct_holder<void, ##__VA_ARGS__> {  \
+    };                                                                                            \
+    template <typename type>                                                                      \
+    class type_caster<holder_type, enable_if_t<!is_shared_ptr<holder_type>::value>>               \
+        : public type_caster_holder<type, holder_type> {};                                        \
+    }                                                                                             \
+    PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)
+
+// PYBIND11_DECLARE_HOLDER_TYPE holder types:
+template <typename base, typename holder>
+struct is_holder_type
+    : std::is_base_of<detail::type_caster_holder<base, holder>, detail::type_caster<holder>> {};
+// Specialization for always-supported unique_ptr holders:
+template <typename base, typename deleter>
+struct is_holder_type<base, std::unique_ptr<base, deleter>> : std::true_type {};
+
+#ifdef PYBIND11_DISABLE_HANDLE_TYPE_NAME_DEFAULT_IMPLEMENTATION // See PR #4888
+
+// This leads to compilation errors if a specialization is missing.
+template <typename T>
+struct handle_type_name;
+
+#else
+
+template <typename T>
+struct handle_type_name {
+    static constexpr auto name = const_name<T>();
+};
+
+#endif
+
+template <>
+struct handle_type_name<object> {
+    static constexpr auto name = const_name("object");
+};
+template <>
+struct handle_type_name<list> {
+    static constexpr auto name = const_name("list");
+};
+template <>
+struct handle_type_name<dict> {
+    static constexpr auto name = const_name("dict");
+};
+template <>
+struct handle_type_name<anyset> {
+    static constexpr auto name = const_name("Union[set, frozenset]");
+};
+template <>
+struct handle_type_name<set> {
+    static constexpr auto name = const_name("set");
+};
+template <>
+struct handle_type_name<frozenset> {
+    static constexpr auto name = const_name("frozenset");
+};
+template <>
+struct handle_type_name<str> {
+    static constexpr auto name = const_name("str");
+};
+template <>
+struct handle_type_name<tuple> {
+    static constexpr auto name = const_name("tuple");
+};
+template <>
+struct handle_type_name<bool_> {
+    static constexpr auto name = const_name("bool");
+};
+template <>
+struct handle_type_name<bytes> {
+    static constexpr auto name = const_name(PYBIND11_BYTES_NAME);
+};
+template <>
+struct handle_type_name<buffer> {
+    static constexpr auto name = const_name("Buffer");
+};
+template <>
+struct handle_type_name<int_> {
+    static constexpr auto name = const_name("int");
+};
+template <>
+struct handle_type_name<iterable> {
+    static constexpr auto name = const_name("Iterable");
+};
+template <>
+struct handle_type_name<iterator> {
+    static constexpr auto name = const_name("Iterator");
+};
+template <>
+struct handle_type_name<float_> {
+    static constexpr auto name = const_name("float");
+};
+template <>
+struct handle_type_name<function> {
+    static constexpr auto name = const_name("Callable");
+};
+template <>
+struct handle_type_name<handle> {
+    static constexpr auto name = handle_type_name<object>::name;
+};
+template <>
+struct handle_type_name<none> {
+    static constexpr auto name = const_name("None");
+};
+template <>
+struct handle_type_name<sequence> {
+    static constexpr auto name = const_name("Sequence");
+};
+template <>
+struct handle_type_name<bytearray> {
+    static constexpr auto name = const_name("bytearray");
+};
+template <>
+struct handle_type_name<memoryview> {
+    static constexpr auto name = const_name("memoryview");
+};
+template <>
+struct handle_type_name<slice> {
+    static constexpr auto name = const_name("slice");
+};
+template <>
+struct handle_type_name<type> {
+    static constexpr auto name = const_name("type");
+};
+template <>
+struct handle_type_name<capsule> {
+    static constexpr auto name = const_name("capsule");
+};
+template <>
+struct handle_type_name<ellipsis> {
+    static constexpr auto name = const_name("ellipsis");
+};
+template <>
+struct handle_type_name<weakref> {
+    static constexpr auto name = const_name("weakref");
+};
+template <>
+struct handle_type_name<args> {
+    static constexpr auto name = const_name("*args");
+};
+template <>
+struct handle_type_name<kwargs> {
+    static constexpr auto name = const_name("**kwargs");
+};
+template <>
+struct handle_type_name<obj_attr_accessor> {
+    static constexpr auto name = const_name<obj_attr_accessor>();
+};
+template <>
+struct handle_type_name<str_attr_accessor> {
+    static constexpr auto name = const_name<str_attr_accessor>();
+};
+template <>
+struct handle_type_name<item_accessor> {
+    static constexpr auto name = const_name<item_accessor>();
+};
+template <>
+struct handle_type_name<sequence_accessor> {
+    static constexpr auto name = const_name<sequence_accessor>();
+};
+template <>
+struct handle_type_name<list_accessor> {
+    static constexpr auto name = const_name<list_accessor>();
+};
+template <>
+struct handle_type_name<tuple_accessor> {
+    static constexpr auto name = const_name<tuple_accessor>();
+};
+
+template <typename type>
+struct pyobject_caster {
+    template <typename T = type, enable_if_t<std::is_same<T, handle>::value, int> = 0>
+    pyobject_caster() : value() {}
+
+    // `type` may not be default constructible (e.g. frozenset, anyset).  Initializing `value`
+    // to a nil handle is safe since it will only be accessed if `load` succeeds.
+    template <typename T = type, enable_if_t<std::is_base_of<object, T>::value, int> = 0>
+    pyobject_caster() : value(reinterpret_steal<type>(handle())) {}
+
+    template <typename T = type, enable_if_t<std::is_same<T, handle>::value, int> = 0>
+    bool load(handle src, bool /* convert */) {
+        value = src;
+        return static_cast<bool>(value);
+    }
+
+    template <typename T = type, enable_if_t<std::is_base_of<object, T>::value, int> = 0>
+    bool load(handle src, bool /* convert */) {
+        if (!isinstance<type>(src)) {
+            return false;
+        }
+        value = reinterpret_borrow<type>(src);
+        return true;
+    }
+
+    static handle cast(const handle &src, return_value_policy /* policy */, handle /* parent */) {
+        return src.inc_ref();
+    }
+    PYBIND11_TYPE_CASTER(type, handle_type_name<type>::name);
+};
+
+template <typename T>
+class type_caster<T, enable_if_t<is_pyobject<T>::value>> : public pyobject_caster<T> {};
+
+// Our conditions for enabling moving are quite restrictive:
+// At compile time:
+// - T needs to be a non-const, non-pointer, non-reference type
+// - type_caster<T>::operator T&() must exist
+// - the type must be move constructible (obviously)
+// At run-time:
+// - if the type is non-copy-constructible, the object must be the sole owner of the type (i.e. it
+//   must have ref_count() == 1)h
+// If any of the above are not satisfied, we fall back to copying.
+template <typename T>
+using move_is_plain_type
+    = satisfies_none_of<T, std::is_void, std::is_pointer, std::is_reference, std::is_const>;
+template <typename T, typename SFINAE = void>
+struct move_always : std::false_type {};
+template <typename T>
+struct move_always<
+    T,
+    enable_if_t<
+        all_of<move_is_plain_type<T>,
+               negation<is_copy_constructible<T>>,
+               is_move_constructible<T>,
+               std::is_same<decltype(std::declval<make_caster<T>>().operator T &()), T &>>::value>>
+    : std::true_type {};
+template <typename T, typename SFINAE = void>
+struct move_if_unreferenced : std::false_type {};
+template <typename T>
+struct move_if_unreferenced<
+    T,
+    enable_if_t<
+        all_of<move_is_plain_type<T>,
+               negation<move_always<T>>,
+               is_move_constructible<T>,
+               std::is_same<decltype(std::declval<make_caster<T>>().operator T &()), T &>>::value>>
+    : std::true_type {};
+template <typename T>
+using move_never = none_of<move_always<T>, move_if_unreferenced<T>>;
+
+// Detect whether returning a `type` from a cast on type's type_caster is going to result in a
+// reference or pointer to a local variable of the type_caster.  Basically, only
+// non-reference/pointer `type`s and reference/pointers from a type_caster_generic are safe;
+// everything else returns a reference/pointer to a local variable.
+template <typename type>
+using cast_is_temporary_value_reference
+    = bool_constant<(std::is_reference<type>::value || std::is_pointer<type>::value)
+                    && !std::is_base_of<type_caster_generic, make_caster<type>>::value
+                    && !std::is_same<intrinsic_t<type>, void>::value>;
+
+// When a value returned from a C++ function is being cast back to Python, we almost always want to
+// force `policy = move`, regardless of the return value policy the function/method was declared
+// with.
+template <typename Return, typename SFINAE = void>
+struct return_value_policy_override {
+    static return_value_policy policy(return_value_policy p) { return p; }
+};
+
+template <typename Return>
+struct return_value_policy_override<
+    Return,
+    detail::enable_if_t<std::is_base_of<type_caster_generic, make_caster<Return>>::value, void>> {
+    static return_value_policy policy(return_value_policy p) {
+        return !std::is_lvalue_reference<Return>::value && !std::is_pointer<Return>::value
+                   ? return_value_policy::move
+                   : p;
+    }
+};
+
+// Basic python -> C++ casting; throws if casting fails
+template <typename T, typename SFINAE>
+type_caster<T, SFINAE> &load_type(type_caster<T, SFINAE> &conv, const handle &handle) {
+    static_assert(!detail::is_pyobject<T>::value,
+                  "Internal error: type_caster should only be used for C++ types");
+    if (!conv.load(handle, true)) {
+#if !defined(PYBIND11_DETAILED_ERROR_MESSAGES)
+        throw cast_error(
+            "Unable to cast Python instance of type "
+            + str(type::handle_of(handle)).cast<std::string>()
+            + " to C++ type '?' (#define "
+              "PYBIND11_DETAILED_ERROR_MESSAGES or compile in debug mode for details)");
+#else
+        throw cast_error("Unable to cast Python instance of type "
+                         + str(type::handle_of(handle)).cast<std::string>() + " to C++ type '"
+                         + type_id<T>() + "'");
+#endif
+    }
+    return conv;
+}
+// Wrapper around the above that also constructs and returns a type_caster
+template <typename T>
+make_caster<T> load_type(const handle &handle) {
+    make_caster<T> conv;
+    load_type(conv, handle);
+    return conv;
+}
+
+PYBIND11_NAMESPACE_END(detail)
+
+// pytype -> C++ type
+template <typename T,
+          detail::enable_if_t<!detail::is_pyobject<T>::value
+                                  && !detail::is_same_ignoring_cvref<T, PyObject *>::value,
+                              int>
+          = 0>
+T cast(const handle &handle) {
+    using namespace detail;
+    static_assert(!cast_is_temporary_value_reference<T>::value,
+                  "Unable to cast type to reference: value is local to type caster");
+    return cast_op<T>(load_type<T>(handle));
+}
+
+// pytype -> pytype (calls converting constructor)
+template <typename T, detail::enable_if_t<detail::is_pyobject<T>::value, int> = 0>
+T cast(const handle &handle) {
+    return T(reinterpret_borrow<object>(handle));
+}
+
+// Note that `cast<PyObject *>(obj)` increments the reference count of `obj`.
+// This is necessary for the case that `obj` is a temporary, and could
+// not possibly be different, given
+// 1. the established convention that the passed `handle` is borrowed, and
+// 2. we don't want to force all generic code using `cast<T>()` to special-case
+//    handling of `T` = `PyObject *` (to increment the reference count there).
+// It is the responsibility of the caller to ensure that the reference count
+// is decremented.
+template <typename T,
+          typename Handle,
+          detail::enable_if_t<detail::is_same_ignoring_cvref<T, PyObject *>::value
+                                  && detail::is_same_ignoring_cvref<Handle, handle>::value,
+                              int>
+          = 0>
+T cast(Handle &&handle) {
+    return handle.inc_ref().ptr();
+}
+// To optimize way an inc_ref/dec_ref cycle:
+template <typename T,
+          typename Object,
+          detail::enable_if_t<detail::is_same_ignoring_cvref<T, PyObject *>::value
+                                  && detail::is_same_ignoring_cvref<Object, object>::value,
+                              int>
+          = 0>
+T cast(Object &&obj) {
+    return obj.release().ptr();
+}
+
+// C++ type -> py::object
+template <typename T, detail::enable_if_t<!detail::is_pyobject<T>::value, int> = 0>
+object cast(T &&value,
+            return_value_policy policy = return_value_policy::automatic_reference,
+            handle parent = handle()) {
+    using no_ref_T = typename std::remove_reference<T>::type;
+    if (policy == return_value_policy::automatic) {
+        policy = std::is_pointer<no_ref_T>::value     ? return_value_policy::take_ownership
+                 : std::is_lvalue_reference<T>::value ? return_value_policy::copy
+                                                      : return_value_policy::move;
+    } else if (policy == return_value_policy::automatic_reference) {
+        policy = std::is_pointer<no_ref_T>::value     ? return_value_policy::reference
+                 : std::is_lvalue_reference<T>::value ? return_value_policy::copy
+                                                      : return_value_policy::move;
+    }
+    return reinterpret_steal<object>(
+        detail::make_caster<T>::cast(std::forward<T>(value), policy, parent));
+}
+
+template <typename T>
+T handle::cast() const {
+    return pybind11::cast<T>(*this);
+}
+template <>
+inline void handle::cast() const {
+    return;
+}
+
+template <typename T>
+detail::enable_if_t<!detail::move_never<T>::value, T> move(object &&obj) {
+    if (obj.ref_count() > 1) {
+#if !defined(PYBIND11_DETAILED_ERROR_MESSAGES)
+        throw cast_error(
+            "Unable to cast Python " + str(type::handle_of(obj)).cast<std::string>()
+            + " instance to C++ rvalue: instance has multiple references"
+              " (#define PYBIND11_DETAILED_ERROR_MESSAGES or compile in debug mode for details)");
+#else
+        throw cast_error("Unable to move from Python "
+                         + str(type::handle_of(obj)).cast<std::string>() + " instance to C++ "
+                         + type_id<T>() + " instance: instance has multiple references");
+#endif
+    }
+
+    // Move into a temporary and return that, because the reference may be a local value of `conv`
+    T ret = std::move(detail::load_type<T>(obj).operator T &());
+    return ret;
+}
+
+// Calling cast() on an rvalue calls pybind11::cast with the object rvalue, which does:
+// - If we have to move (because T has no copy constructor), do it.  This will fail if the moved
+//   object has multiple references, but trying to copy will fail to compile.
+// - If both movable and copyable, check ref count: if 1, move; otherwise copy
+// - Otherwise (not movable), copy.
+template <typename T>
+detail::enable_if_t<!detail::is_pyobject<T>::value && detail::move_always<T>::value, T>
+cast(object &&object) {
+    return move<T>(std::move(object));
+}
+template <typename T>
+detail::enable_if_t<!detail::is_pyobject<T>::value && detail::move_if_unreferenced<T>::value, T>
+cast(object &&object) {
+    if (object.ref_count() > 1) {
+        return cast<T>(object);
+    }
+    return move<T>(std::move(object));
+}
+template <typename T>
+detail::enable_if_t<!detail::is_pyobject<T>::value && detail::move_never<T>::value, T>
+cast(object &&object) {
+    return cast<T>(object);
+}
+
+// pytype rvalue -> pytype (calls converting constructor)
+template <typename T>
+detail::enable_if_t<detail::is_pyobject<T>::value, T> cast(object &&object) {
+    return T(std::move(object));
+}
+
+template <typename T>
+T object::cast() const & {
+    return pybind11::cast<T>(*this);
+}
+template <typename T>
+T object::cast() && {
+    return pybind11::cast<T>(std::move(*this));
+}
+template <>
+inline void object::cast() const & {
+    return;
+}
+template <>
+inline void object::cast() && {
+    return;
+}
+
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+// Declared in pytypes.h:
+template <typename T, enable_if_t<!is_pyobject<T>::value, int>>
+object object_or_cast(T &&o) {
+    return pybind11::cast(std::forward<T>(o));
+}
+
+// Placeholder type for the unneeded (and dead code) static variable in the
+// PYBIND11_OVERRIDE_OVERRIDE macro
+struct override_unused {};
+template <typename ret_type>
+using override_caster_t = conditional_t<cast_is_temporary_value_reference<ret_type>::value,
+                                        make_caster<ret_type>,
+                                        override_unused>;
+
+// Trampoline use: for reference/pointer types to value-converted values, we do a value cast, then
+// store the result in the given variable.  For other types, this is a no-op.
+template <typename T>
+enable_if_t<cast_is_temporary_value_reference<T>::value, T> cast_ref(object &&o,
+                                                                     make_caster<T> &caster) {
+    return cast_op<T>(load_type(caster, o));
+}
+template <typename T>
+enable_if_t<!cast_is_temporary_value_reference<T>::value, T> cast_ref(object &&,
+                                                                      override_unused &) {
+    pybind11_fail("Internal error: cast_ref fallback invoked");
+}
+
+// Trampoline use: Having a pybind11::cast with an invalid reference type is going to
+// static_assert, even though if it's in dead code, so we provide a "trampoline" to pybind11::cast
+// that only does anything in cases where pybind11::cast is valid.
+template <typename T>
+enable_if_t<cast_is_temporary_value_reference<T>::value
+                && !detail::is_same_ignoring_cvref<T, PyObject *>::value,
+            T>
+cast_safe(object &&) {
+    pybind11_fail("Internal error: cast_safe fallback invoked");
+}
+template <typename T>
+enable_if_t<std::is_void<T>::value, void> cast_safe(object &&) {}
+template <typename T>
+enable_if_t<detail::is_same_ignoring_cvref<T, PyObject *>::value, PyObject *>
+cast_safe(object &&o) {
+    return o.release().ptr();
+}
+template <typename T>
+enable_if_t<detail::none_of<cast_is_temporary_value_reference<T>,
+                            detail::is_same_ignoring_cvref<T, PyObject *>,
+                            std::is_void<T>>::value,
+            T>
+cast_safe(object &&o) {
+    return pybind11::cast<T>(std::move(o));
+}
+
+PYBIND11_NAMESPACE_END(detail)
+
+// The overloads could coexist, i.e. the #if is not strictly speaking needed,
+// but it is an easy minor optimization.
+#if !defined(PYBIND11_DETAILED_ERROR_MESSAGES)
+inline cast_error cast_error_unable_to_convert_call_arg(const std::string &name) {
+    return cast_error("Unable to convert call argument '" + name
+                      + "' to Python object (#define "
+                        "PYBIND11_DETAILED_ERROR_MESSAGES or compile in debug mode for details)");
+}
+#else
+inline cast_error cast_error_unable_to_convert_call_arg(const std::string &name,
+                                                        const std::string &type) {
+    return cast_error("Unable to convert call argument '" + name + "' of type '" + type
+                      + "' to Python object");
+}
+#endif
+
+template <return_value_policy policy = return_value_policy::automatic_reference>
+tuple make_tuple() {
+    return tuple(0);
+}
+
+template <return_value_policy policy = return_value_policy::automatic_reference, typename... Args>
+tuple make_tuple(Args &&...args_) {
+    constexpr size_t size = sizeof...(Args);
+    std::array<object, size> args{{reinterpret_steal<object>(
+        detail::make_caster<Args>::cast(std::forward<Args>(args_), policy, nullptr))...}};
+    for (size_t i = 0; i < args.size(); i++) {
+        if (!args[i]) {
+#if !defined(PYBIND11_DETAILED_ERROR_MESSAGES)
+            throw cast_error_unable_to_convert_call_arg(std::to_string(i));
+#else
+            std::array<std::string, size> argtypes{{type_id<Args>()...}};
+            throw cast_error_unable_to_convert_call_arg(std::to_string(i), argtypes[i]);
+#endif
+        }
+    }
+    tuple result(size);
+    int counter = 0;
+    for (auto &arg_value : args) {
+        PyTuple_SET_ITEM(result.ptr(), counter++, arg_value.release().ptr());
+    }
+    return result;
+}
+
+/// \ingroup annotations
+/// Annotation for arguments
+struct arg {
+    /// Constructs an argument with the name of the argument; if null or omitted, this is a
+    /// positional argument.
+    constexpr explicit arg(const char *name = nullptr)
+        : name(name), flag_noconvert(false), flag_none(true) {}
+    /// Assign a value to this argument
+    template <typename T>
+    arg_v operator=(T &&value) const;
+    /// Indicate that the type should not be converted in the type caster
+    arg &noconvert(bool flag = true) {
+        flag_noconvert = flag;
+        return *this;
+    }
+    /// Indicates that the argument should/shouldn't allow None (e.g. for nullable pointer args)
+    arg &none(bool flag = true) {
+        flag_none = flag;
+        return *this;
+    }
+
+    const char *name;        ///< If non-null, this is a named kwargs argument
+    bool flag_noconvert : 1; ///< If set, do not allow conversion (requires a supporting type
+                             ///< caster!)
+    bool flag_none : 1;      ///< If set (the default), allow None to be passed to this argument
+};
+
+/// \ingroup annotations
+/// Annotation for arguments with values
+struct arg_v : arg {
+private:
+    template <typename T>
+    arg_v(arg &&base, T &&x, const char *descr = nullptr)
+        : arg(base), value(reinterpret_steal<object>(detail::make_caster<T>::cast(
+                         std::forward<T>(x), return_value_policy::automatic, {}))),
+          descr(descr)
+#if defined(PYBIND11_DETAILED_ERROR_MESSAGES)
+          ,
+          type(type_id<T>())
+#endif
+    {
+        // Workaround! See:
+        // https://github.com/pybind/pybind11/issues/2336
+        // https://github.com/pybind/pybind11/pull/2685#issuecomment-731286700
+        if (PyErr_Occurred()) {
+            PyErr_Clear();
+        }
+    }
+
+public:
+    /// Direct construction with name, default, and description
+    template <typename T>
+    arg_v(const char *name, T &&x, const char *descr = nullptr)
+        : arg_v(arg(name), std::forward<T>(x), descr) {}
+
+    /// Called internally when invoking `py::arg("a") = value`
+    template <typename T>
+    arg_v(const arg &base, T &&x, const char *descr = nullptr)
+        : arg_v(arg(base), std::forward<T>(x), descr) {}
+
+    /// Same as `arg::noconvert()`, but returns *this as arg_v&, not arg&
+    arg_v &noconvert(bool flag = true) {
+        arg::noconvert(flag);
+        return *this;
+    }
+
+    /// Same as `arg::nonone()`, but returns *this as arg_v&, not arg&
+    arg_v &none(bool flag = true) {
+        arg::none(flag);
+        return *this;
+    }
+
+    /// The default value
+    object value;
+    /// The (optional) description of the default value
+    const char *descr;
+#if defined(PYBIND11_DETAILED_ERROR_MESSAGES)
+    /// The C++ type name of the default value (only available when compiled in debug mode)
+    std::string type;
+#endif
+};
+
+/// \ingroup annotations
+/// Annotation indicating that all following arguments are keyword-only; the is the equivalent of
+/// an unnamed '*' argument
+struct kw_only {};
+
+/// \ingroup annotations
+/// Annotation indicating that all previous arguments are positional-only; the is the equivalent of
+/// an unnamed '/' argument (in Python 3.8)
+struct pos_only {};
+
+template <typename T>
+arg_v arg::operator=(T &&value) const {
+    return {*this, std::forward<T>(value)};
+}
+
+/// Alias for backward compatibility -- to be removed in version 2.0
+template <typename /*unused*/>
+using arg_t = arg_v;
+
+inline namespace literals {
+/** \rst
+    String literal version of `arg`
+ \endrst */
+constexpr arg
+#if !defined(__clang__) && defined(__GNUC__) && __GNUC__ < 5
+operator"" _a // gcc 4.8.5 insists on having a space (hard error).
+#else
+operator""_a // clang 17 generates a deprecation warning if there is a space.
+#endif
+    (const char *name, size_t) {
+    return arg(name);
+}
+} // namespace literals
+
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+template <typename T>
+using is_kw_only = std::is_same<intrinsic_t<T>, kw_only>;
+template <typename T>
+using is_pos_only = std::is_same<intrinsic_t<T>, pos_only>;
+
+// forward declaration (definition in attr.h)
+struct function_record;
+
+/// Internal data associated with a single function call
+struct function_call {
+    function_call(const function_record &f, handle p); // Implementation in attr.h
+
+    /// The function data:
+    const function_record &func;
+
+    /// Arguments passed to the function:
+    std::vector<handle> args;
+
+    /// The `convert` value the arguments should be loaded with
+    std::vector<bool> args_convert;
+
+    /// Extra references for the optional `py::args` and/or `py::kwargs` arguments (which, if
+    /// present, are also in `args` but without a reference).
+    object args_ref, kwargs_ref;
+
+    /// The parent, if any
+    handle parent;
+
+    /// If this is a call to an initializer, this argument contains `self`
+    handle init_self;
+};
+
+/// Helper class which loads arguments for C++ functions called from Python
+template <typename... Args>
+class argument_loader {
+    using indices = make_index_sequence<sizeof...(Args)>;
+
+    template <typename Arg>
+    using argument_is_args = std::is_same<intrinsic_t<Arg>, args>;
+    template <typename Arg>
+    using argument_is_kwargs = std::is_same<intrinsic_t<Arg>, kwargs>;
+    // Get kwargs argument position, or -1 if not present:
+    static constexpr auto kwargs_pos = constexpr_last<argument_is_kwargs, Args...>();
+
+    static_assert(kwargs_pos == -1 || kwargs_pos == (int) sizeof...(Args) - 1,
+                  "py::kwargs is only permitted as the last argument of a function");
+
+public:
+    static constexpr bool has_kwargs = kwargs_pos != -1;
+
+    // py::args argument position; -1 if not present.
+    static constexpr int args_pos = constexpr_last<argument_is_args, Args...>();
+
+    static_assert(args_pos == -1 || args_pos == constexpr_first<argument_is_args, Args...>(),
+                  "py::args cannot be specified more than once");
+
+    static constexpr auto arg_names
+        = ::pybind11::detail::concat(type_descr(make_caster<Args>::name)...);
+
+    bool load_args(function_call &call) { return load_impl_sequence(call, indices{}); }
+
+    template <typename Return, typename Guard, typename Func>
+    // NOLINTNEXTLINE(readability-const-return-type)
+    enable_if_t<!std::is_void<Return>::value, Return> call(Func &&f) && {
+        return std::move(*this).template call_impl<remove_cv_t<Return>>(
+            std::forward<Func>(f), indices{}, Guard{});
+    }
+
+    template <typename Return, typename Guard, typename Func>
+    enable_if_t<std::is_void<Return>::value, void_type> call(Func &&f) && {
+        std::move(*this).template call_impl<remove_cv_t<Return>>(
+            std::forward<Func>(f), indices{}, Guard{});
+        return void_type();
+    }
+
+private:
+    static bool load_impl_sequence(function_call &, index_sequence<>) { return true; }
+
+    template <size_t... Is>
+    bool load_impl_sequence(function_call &call, index_sequence<Is...>) {
+#ifdef __cpp_fold_expressions
+        if ((... || !std::get<Is>(argcasters).load(call.args[Is], call.args_convert[Is]))) {
+            return false;
+        }
+#else
+        for (bool r : {std::get<Is>(argcasters).load(call.args[Is], call.args_convert[Is])...}) {
+            if (!r) {
+                return false;
+            }
+        }
+#endif
+        return true;
+    }
+
+    template <typename Return, typename Func, size_t... Is, typename Guard>
+    Return call_impl(Func &&f, index_sequence<Is...>, Guard &&) && {
+        return std::forward<Func>(f)(cast_op<Args>(std::move(std::get<Is>(argcasters)))...);
+    }
+
+    std::tuple<make_caster<Args>...> argcasters;
+};
+
+/// Helper class which collects only positional arguments for a Python function call.
+/// A fancier version below can collect any argument, but this one is optimal for simple calls.
+template <return_value_policy policy>
+class simple_collector {
+public:
+    template <typename... Ts>
+    explicit simple_collector(Ts &&...values)
+        : m_args(pybind11::make_tuple<policy>(std::forward<Ts>(values)...)) {}
+
+    const tuple &args() const & { return m_args; }
+    dict kwargs() const { return {}; }
+
+    tuple args() && { return std::move(m_args); }
+
+    /// Call a Python function and pass the collected arguments
+    object call(PyObject *ptr) const {
+        PyObject *result = PyObject_CallObject(ptr, m_args.ptr());
+        if (!result) {
+            throw error_already_set();
+        }
+        return reinterpret_steal<object>(result);
+    }
+
+private:
+    tuple m_args;
+};
+
+/// Helper class which collects positional, keyword, * and ** arguments for a Python function call
+template <return_value_policy policy>
+class unpacking_collector {
+public:
+    template <typename... Ts>
+    explicit unpacking_collector(Ts &&...values) {
+        // Tuples aren't (easily) resizable so a list is needed for collection,
+        // but the actual function call strictly requires a tuple.
+        auto args_list = list();
+        using expander = int[];
+        (void) expander{0, (process(args_list, std::forward<Ts>(values)), 0)...};
+
+        m_args = std::move(args_list);
+    }
+
+    const tuple &args() const & { return m_args; }
+    const dict &kwargs() const & { return m_kwargs; }
+
+    tuple args() && { return std::move(m_args); }
+    dict kwargs() && { return std::move(m_kwargs); }
+
+    /// Call a Python function and pass the collected arguments
+    object call(PyObject *ptr) const {
+        PyObject *result = PyObject_Call(ptr, m_args.ptr(), m_kwargs.ptr());
+        if (!result) {
+            throw error_already_set();
+        }
+        return reinterpret_steal<object>(result);
+    }
+
+private:
+    template <typename T>
+    void process(list &args_list, T &&x) {
+        auto o = reinterpret_steal<object>(
+            detail::make_caster<T>::cast(std::forward<T>(x), policy, {}));
+        if (!o) {
+#if !defined(PYBIND11_DETAILED_ERROR_MESSAGES)
+            throw cast_error_unable_to_convert_call_arg(std::to_string(args_list.size()));
+#else
+            throw cast_error_unable_to_convert_call_arg(std::to_string(args_list.size()),
+                                                        type_id<T>());
+#endif
+        }
+        args_list.append(std::move(o));
+    }
+
+    void process(list &args_list, detail::args_proxy ap) {
+        for (auto a : ap) {
+            args_list.append(a);
+        }
+    }
+
+    void process(list & /*args_list*/, arg_v a) {
+        if (!a.name) {
+#if !defined(PYBIND11_DETAILED_ERROR_MESSAGES)
+            nameless_argument_error();
+#else
+            nameless_argument_error(a.type);
+#endif
+        }
+        if (m_kwargs.contains(a.name)) {
+#if !defined(PYBIND11_DETAILED_ERROR_MESSAGES)
+            multiple_values_error();
+#else
+            multiple_values_error(a.name);
+#endif
+        }
+        if (!a.value) {
+#if !defined(PYBIND11_DETAILED_ERROR_MESSAGES)
+            throw cast_error_unable_to_convert_call_arg(a.name);
+#else
+            throw cast_error_unable_to_convert_call_arg(a.name, a.type);
+#endif
+        }
+        m_kwargs[a.name] = std::move(a.value);
+    }
+
+    void process(list & /*args_list*/, detail::kwargs_proxy kp) {
+        if (!kp) {
+            return;
+        }
+        for (auto k : reinterpret_borrow<dict>(kp)) {
+            if (m_kwargs.contains(k.first)) {
+#if !defined(PYBIND11_DETAILED_ERROR_MESSAGES)
+                multiple_values_error();
+#else
+                multiple_values_error(str(k.first));
+#endif
+            }
+            m_kwargs[k.first] = k.second;
+        }
+    }
+
+    [[noreturn]] static void nameless_argument_error() {
+        throw type_error(
+            "Got kwargs without a name; only named arguments "
+            "may be passed via py::arg() to a python function call. "
+            "(#define PYBIND11_DETAILED_ERROR_MESSAGES or compile in debug mode for details)");
+    }
+    [[noreturn]] static void nameless_argument_error(const std::string &type) {
+        throw type_error("Got kwargs without a name of type '" + type
+                         + "'; only named "
+                           "arguments may be passed via py::arg() to a python function call. ");
+    }
+    [[noreturn]] static void multiple_values_error() {
+        throw type_error(
+            "Got multiple values for keyword argument "
+            "(#define PYBIND11_DETAILED_ERROR_MESSAGES or compile in debug mode for details)");
+    }
+
+    [[noreturn]] static void multiple_values_error(const std::string &name) {
+        throw type_error("Got multiple values for keyword argument '" + name + "'");
+    }
+
+private:
+    tuple m_args;
+    dict m_kwargs;
+};
+
+// [workaround(intel)] Separate function required here
+// We need to put this into a separate function because the Intel compiler
+// fails to compile enable_if_t<!all_of<is_positional<Args>...>::value>
+// (tested with ICC 2021.1 Beta 20200827).
+template <typename... Args>
+constexpr bool args_are_all_positional() {
+    return all_of<is_positional<Args>...>::value;
+}
+
+/// Collect only positional arguments for a Python function call
+template <return_value_policy policy,
+          typename... Args,
+          typename = enable_if_t<args_are_all_positional<Args...>()>>
+simple_collector<policy> collect_arguments(Args &&...args) {
+    return simple_collector<policy>(std::forward<Args>(args)...);
+}
+
+/// Collect all arguments, including keywords and unpacking (only instantiated when needed)
+template <return_value_policy policy,
+          typename... Args,
+          typename = enable_if_t<!args_are_all_positional<Args...>()>>
+unpacking_collector<policy> collect_arguments(Args &&...args) {
+    // Following argument order rules for generalized unpacking according to PEP 448
+    static_assert(constexpr_last<is_positional, Args...>()
+                          < constexpr_first<is_keyword_or_ds, Args...>()
+                      && constexpr_last<is_s_unpacking, Args...>()
+                             < constexpr_first<is_ds_unpacking, Args...>(),
+                  "Invalid function call: positional args must precede keywords and ** unpacking; "
+                  "* unpacking must precede ** unpacking");
+    return unpacking_collector<policy>(std::forward<Args>(args)...);
+}
+
+template <typename Derived>
+template <return_value_policy policy, typename... Args>
+object object_api<Derived>::operator()(Args &&...args) const {
+#ifndef NDEBUG
+    if (!PyGILState_Check()) {
+        pybind11_fail("pybind11::object_api<>::operator() PyGILState_Check() failure.");
+    }
+#endif
+    return detail::collect_arguments<policy>(std::forward<Args>(args)...).call(derived().ptr());
+}
+
+template <typename Derived>
+template <return_value_policy policy, typename... Args>
+object object_api<Derived>::call(Args &&...args) const {
+    return operator()<policy>(std::forward<Args>(args)...);
+}
+
+PYBIND11_NAMESPACE_END(detail)
+
+template <typename T>
+handle type::handle_of() {
+    static_assert(std::is_base_of<detail::type_caster_generic, detail::make_caster<T>>::value,
+                  "py::type::of<T> only supports the case where T is a registered C++ types.");
+
+    return detail::get_type_handle(typeid(T), true);
+}
+
+#define PYBIND11_MAKE_OPAQUE(...)                                                                 \
+    PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)                                                  \
+    namespace detail {                                                                            \
+    template <>                                                                                   \
+    class type_caster<__VA_ARGS__> : public type_caster_base<__VA_ARGS__> {};                     \
+    }                                                                                             \
+    PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)
+
+/// Lets you pass a type containing a `,` through a macro parameter without needing a separate
+/// typedef, e.g.:
+/// `PYBIND11_OVERRIDE(PYBIND11_TYPE(ReturnType<A, B>), PYBIND11_TYPE(Parent<C, D>), f, arg)`
+#define PYBIND11_TYPE(...) __VA_ARGS__
+
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/chrono.h

@@ -0,0 +1,225 @@
+/*
+    pybind11/chrono.h: Transparent conversion between std::chrono and python's datetime
+
+    Copyright (c) 2016 Trent Houliston <trent@houliston.me> and
+                       Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "pybind11.h"
+
+#include <chrono>
+#include <cmath>
+#include <ctime>
+#include <datetime.h>
+#include <mutex>
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+template <typename type>
+class duration_caster {
+public:
+    using rep = typename type::rep;
+    using period = typename type::period;
+
+    // signed 25 bits required by the standard.
+    using days = std::chrono::duration<int_least32_t, std::ratio<86400>>;
+
+    bool load(handle src, bool) {
+        using namespace std::chrono;
+
+        // Lazy initialise the PyDateTime import
+        if (!PyDateTimeAPI) {
+            PyDateTime_IMPORT;
+        }
+
+        if (!src) {
+            return false;
+        }
+        // If invoked with datetime.delta object
+        if (PyDelta_Check(src.ptr())) {
+            value = type(duration_cast<duration<rep, period>>(
+                days(PyDateTime_DELTA_GET_DAYS(src.ptr()))
+                + seconds(PyDateTime_DELTA_GET_SECONDS(src.ptr()))
+                + microseconds(PyDateTime_DELTA_GET_MICROSECONDS(src.ptr()))));
+            return true;
+        }
+        // If invoked with a float we assume it is seconds and convert
+        if (PyFloat_Check(src.ptr())) {
+            value = type(duration_cast<duration<rep, period>>(
+                duration<double>(PyFloat_AsDouble(src.ptr()))));
+            return true;
+        }
+        return false;
+    }
+
+    // If this is a duration just return it back
+    static const std::chrono::duration<rep, period> &
+    get_duration(const std::chrono::duration<rep, period> &src) {
+        return src;
+    }
+
+    // If this is a time_point get the time_since_epoch
+    template <typename Clock>
+    static std::chrono::duration<rep, period>
+    get_duration(const std::chrono::time_point<Clock, std::chrono::duration<rep, period>> &src) {
+        return src.time_since_epoch();
+    }
+
+    static handle cast(const type &src, return_value_policy /* policy */, handle /* parent */) {
+        using namespace std::chrono;
+
+        // Use overloaded function to get our duration from our source
+        // Works out if it is a duration or time_point and get the duration
+        auto d = get_duration(src);
+
+        // Lazy initialise the PyDateTime import
+        if (!PyDateTimeAPI) {
+            PyDateTime_IMPORT;
+        }
+
+        // Declare these special duration types so the conversions happen with the correct
+        // primitive types (int)
+        using dd_t = duration<int, std::ratio<86400>>;
+        using ss_t = duration<int, std::ratio<1>>;
+        using us_t = duration<int, std::micro>;
+
+        auto dd = duration_cast<dd_t>(d);
+        auto subd = d - dd;
+        auto ss = duration_cast<ss_t>(subd);
+        auto us = duration_cast<us_t>(subd - ss);
+        return PyDelta_FromDSU(dd.count(), ss.count(), us.count());
+    }
+
+    PYBIND11_TYPE_CASTER(type, const_name("datetime.timedelta"));
+};
+
+inline std::tm *localtime_thread_safe(const std::time_t *time, std::tm *buf) {
+#if (defined(__STDC_LIB_EXT1__) && defined(__STDC_WANT_LIB_EXT1__)) || defined(_MSC_VER)
+    if (localtime_s(buf, time))
+        return nullptr;
+    return buf;
+#else
+    static std::mutex mtx;
+    std::lock_guard<std::mutex> lock(mtx);
+    std::tm *tm_ptr = std::localtime(time);
+    if (tm_ptr != nullptr) {
+        *buf = *tm_ptr;
+    }
+    return tm_ptr;
+#endif
+}
+
+// This is for casting times on the system clock into datetime.datetime instances
+template <typename Duration>
+class type_caster<std::chrono::time_point<std::chrono::system_clock, Duration>> {
+public:
+    using type = std::chrono::time_point<std::chrono::system_clock, Duration>;
+    bool load(handle src, bool) {
+        using namespace std::chrono;
+
+        // Lazy initialise the PyDateTime import
+        if (!PyDateTimeAPI) {
+            PyDateTime_IMPORT;
+        }
+
+        if (!src) {
+            return false;
+        }
+
+        std::tm cal;
+        microseconds msecs;
+
+        if (PyDateTime_Check(src.ptr())) {
+            cal.tm_sec = PyDateTime_DATE_GET_SECOND(src.ptr());
+            cal.tm_min = PyDateTime_DATE_GET_MINUTE(src.ptr());
+            cal.tm_hour = PyDateTime_DATE_GET_HOUR(src.ptr());
+            cal.tm_mday = PyDateTime_GET_DAY(src.ptr());
+            cal.tm_mon = PyDateTime_GET_MONTH(src.ptr()) - 1;
+            cal.tm_year = PyDateTime_GET_YEAR(src.ptr()) - 1900;
+            cal.tm_isdst = -1;
+            msecs = microseconds(PyDateTime_DATE_GET_MICROSECOND(src.ptr()));
+        } else if (PyDate_Check(src.ptr())) {
+            cal.tm_sec = 0;
+            cal.tm_min = 0;
+            cal.tm_hour = 0;
+            cal.tm_mday = PyDateTime_GET_DAY(src.ptr());
+            cal.tm_mon = PyDateTime_GET_MONTH(src.ptr()) - 1;
+            cal.tm_year = PyDateTime_GET_YEAR(src.ptr()) - 1900;
+            cal.tm_isdst = -1;
+            msecs = microseconds(0);
+        } else if (PyTime_Check(src.ptr())) {
+            cal.tm_sec = PyDateTime_TIME_GET_SECOND(src.ptr());
+            cal.tm_min = PyDateTime_TIME_GET_MINUTE(src.ptr());
+            cal.tm_hour = PyDateTime_TIME_GET_HOUR(src.ptr());
+            cal.tm_mday = 1;  // This date (day, month, year) = (1, 0, 70)
+            cal.tm_mon = 0;   // represents 1-Jan-1970, which is the first
+            cal.tm_year = 70; // earliest available date for Python's datetime
+            cal.tm_isdst = -1;
+            msecs = microseconds(PyDateTime_TIME_GET_MICROSECOND(src.ptr()));
+        } else {
+            return false;
+        }
+
+        value = time_point_cast<Duration>(system_clock::from_time_t(std::mktime(&cal)) + msecs);
+        return true;
+    }
+
+    static handle cast(const std::chrono::time_point<std::chrono::system_clock, Duration> &src,
+                       return_value_policy /* policy */,
+                       handle /* parent */) {
+        using namespace std::chrono;
+
+        // Lazy initialise the PyDateTime import
+        if (!PyDateTimeAPI) {
+            PyDateTime_IMPORT;
+        }
+
+        // Get out microseconds, and make sure they are positive, to avoid bug in eastern
+        // hemisphere time zones (cfr. https://github.com/pybind/pybind11/issues/2417)
+        using us_t = duration<int, std::micro>;
+        auto us = duration_cast<us_t>(src.time_since_epoch() % seconds(1));
+        if (us.count() < 0) {
+            us += seconds(1);
+        }
+
+        // Subtract microseconds BEFORE `system_clock::to_time_t`, because:
+        // > If std::time_t has lower precision, it is implementation-defined whether the value is
+        // rounded or truncated. (https://en.cppreference.com/w/cpp/chrono/system_clock/to_time_t)
+        std::time_t tt
+            = system_clock::to_time_t(time_point_cast<system_clock::duration>(src - us));
+
+        std::tm localtime;
+        std::tm *localtime_ptr = localtime_thread_safe(&tt, &localtime);
+        if (!localtime_ptr) {
+            throw cast_error("Unable to represent system_clock in local time");
+        }
+        return PyDateTime_FromDateAndTime(localtime.tm_year + 1900,
+                                          localtime.tm_mon + 1,
+                                          localtime.tm_mday,
+                                          localtime.tm_hour,
+                                          localtime.tm_min,
+                                          localtime.tm_sec,
+                                          us.count());
+    }
+    PYBIND11_TYPE_CASTER(type, const_name("datetime.datetime"));
+};
+
+// Other clocks that are not the system clock are not measured as datetime.datetime objects
+// since they are not measured on calendar time. So instead we just make them timedeltas
+// Or if they have passed us a time as a float we convert that
+template <typename Clock, typename Duration>
+class type_caster<std::chrono::time_point<Clock, Duration>>
+    : public duration_caster<std::chrono::time_point<Clock, Duration>> {};
+
+template <typename Rep, typename Period>
+class type_caster<std::chrono::duration<Rep, Period>>
+    : public duration_caster<std::chrono::duration<Rep, Period>> {};
+
+PYBIND11_NAMESPACE_END(detail)
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/common.h

@@ -0,0 +1,2 @@
+#include "detail/common.h"
+#warning "Including 'common.h' is deprecated. It will be removed in v3.0. Use 'pybind11.h'."

phivenv/Lib/site-packages/torch/include/pybind11/complex.h

@@ -0,0 +1,74 @@
+/*
+    pybind11/complex.h: Complex number support
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "pybind11.h"
+
+#include <complex>
+
+/// glibc defines I as a macro which breaks things, e.g., boost template names
+#ifdef I
+#    undef I
+#endif
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+template <typename T>
+struct format_descriptor<std::complex<T>, detail::enable_if_t<std::is_floating_point<T>::value>> {
+    static constexpr const char c = format_descriptor<T>::c;
+    static constexpr const char value[3] = {'Z', c, '\0'};
+    static std::string format() { return std::string(value); }
+};
+
+#ifndef PYBIND11_CPP17
+
+template <typename T>
+constexpr const char
+    format_descriptor<std::complex<T>,
+                      detail::enable_if_t<std::is_floating_point<T>::value>>::value[3];
+
+#endif
+
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+template <typename T>
+struct is_fmt_numeric<std::complex<T>, detail::enable_if_t<std::is_floating_point<T>::value>> {
+    static constexpr bool value = true;
+    static constexpr int index = is_fmt_numeric<T>::index + 3;
+};
+
+template <typename T>
+class type_caster<std::complex<T>> {
+public:
+    bool load(handle src, bool convert) {
+        if (!src) {
+            return false;
+        }
+        if (!convert && !PyComplex_Check(src.ptr())) {
+            return false;
+        }
+        Py_complex result = PyComplex_AsCComplex(src.ptr());
+        if (result.real == -1.0 && PyErr_Occurred()) {
+            PyErr_Clear();
+            return false;
+        }
+        value = std::complex<T>((T) result.real, (T) result.imag);
+        return true;
+    }
+
+    static handle
+    cast(const std::complex<T> &src, return_value_policy /* policy */, handle /* parent */) {
+        return PyComplex_FromDoubles((double) src.real(), (double) src.imag());
+    }
+
+    PYBIND11_TYPE_CASTER(std::complex<T>, const_name("complex"));
+};
+PYBIND11_NAMESPACE_END(detail)
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/detail/class.h

@@ -0,0 +1,767 @@
+/*
+    pybind11/detail/class.h: Python C API implementation details for py::class_
+
+    Copyright (c) 2017 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include <pybind11/attr.h>
+#include <pybind11/options.h>
+
+#include "exception_translation.h"
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+#if !defined(PYPY_VERSION)
+#    define PYBIND11_BUILTIN_QUALNAME
+#    define PYBIND11_SET_OLDPY_QUALNAME(obj, nameobj)
+#else
+// In PyPy, we still set __qualname__ so that we can produce reliable function type
+// signatures; in CPython this macro expands to nothing:
+#    define PYBIND11_SET_OLDPY_QUALNAME(obj, nameobj)                                             \
+        setattr((PyObject *) obj, "__qualname__", nameobj)
+#endif
+
+inline std::string get_fully_qualified_tp_name(PyTypeObject *type) {
+#if !defined(PYPY_VERSION)
+    return type->tp_name;
+#else
+    auto module_name = handle((PyObject *) type).attr("__module__").cast<std::string>();
+    if (module_name == PYBIND11_BUILTINS_MODULE)
+        return type->tp_name;
+    else
+        return std::move(module_name) + "." + type->tp_name;
+#endif
+}
+
+inline PyTypeObject *type_incref(PyTypeObject *type) {
+    Py_INCREF(type);
+    return type;
+}
+
+#if !defined(PYPY_VERSION)
+
+/// `pybind11_static_property.__get__()`: Always pass the class instead of the instance.
+extern "C" inline PyObject *pybind11_static_get(PyObject *self, PyObject * /*ob*/, PyObject *cls) {
+    return PyProperty_Type.tp_descr_get(self, cls, cls);
+}
+
+/// `pybind11_static_property.__set__()`: Just like the above `__get__()`.
+extern "C" inline int pybind11_static_set(PyObject *self, PyObject *obj, PyObject *value) {
+    PyObject *cls = PyType_Check(obj) ? obj : (PyObject *) Py_TYPE(obj);
+    return PyProperty_Type.tp_descr_set(self, cls, value);
+}
+
+// Forward declaration to use in `make_static_property_type()`
+inline void enable_dynamic_attributes(PyHeapTypeObject *heap_type);
+
+/** A `static_property` is the same as a `property` but the `__get__()` and `__set__()`
+    methods are modified to always use the object type instead of a concrete instance.
+    Return value: New reference. */
+inline PyTypeObject *make_static_property_type() {
+    constexpr auto *name = "pybind11_static_property";
+    auto name_obj = reinterpret_steal<object>(PYBIND11_FROM_STRING(name));
+
+    /* Danger zone: from now (and until PyType_Ready), make sure to
+       issue no Python C API calls which could potentially invoke the
+       garbage collector (the GC will call type_traverse(), which will in
+       turn find the newly constructed type in an invalid state) */
+    auto *heap_type = (PyHeapTypeObject *) PyType_Type.tp_alloc(&PyType_Type, 0);
+    if (!heap_type) {
+        pybind11_fail("make_static_property_type(): error allocating type!");
+    }
+
+    heap_type->ht_name = name_obj.inc_ref().ptr();
+#    ifdef PYBIND11_BUILTIN_QUALNAME
+    heap_type->ht_qualname = name_obj.inc_ref().ptr();
+#    endif
+
+    auto *type = &heap_type->ht_type;
+    type->tp_name = name;
+    type->tp_base = type_incref(&PyProperty_Type);
+    type->tp_flags = Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE | Py_TPFLAGS_HEAPTYPE;
+    type->tp_descr_get = pybind11_static_get;
+    type->tp_descr_set = pybind11_static_set;
+
+#    if PY_VERSION_HEX >= 0x030C0000
+    // Since Python-3.12 property-derived types are required to
+    // have dynamic attributes (to set `__doc__`)
+    enable_dynamic_attributes(heap_type);
+#    endif
+
+    if (PyType_Ready(type) < 0) {
+        pybind11_fail("make_static_property_type(): failure in PyType_Ready()!");
+    }
+
+    setattr((PyObject *) type, "__module__", str("pybind11_builtins"));
+    PYBIND11_SET_OLDPY_QUALNAME(type, name_obj);
+
+    return type;
+}
+
+#else // PYPY
+
+/** PyPy has some issues with the above C API, so we evaluate Python code instead.
+    This function will only be called once so performance isn't really a concern.
+    Return value: New reference. */
+inline PyTypeObject *make_static_property_type() {
+    auto d = dict();
+    PyObject *result = PyRun_String(R"(\
+class pybind11_static_property(property):
+    def __get__(self, obj, cls):
+        return property.__get__(self, cls, cls)
+
+    def __set__(self, obj, value):
+        cls = obj if isinstance(obj, type) else type(obj)
+        property.__set__(self, cls, value)
+)",
+                                    Py_file_input,
+                                    d.ptr(),
+                                    d.ptr());
+    if (result == nullptr)
+        throw error_already_set();
+    Py_DECREF(result);
+    return (PyTypeObject *) d["pybind11_static_property"].cast<object>().release().ptr();
+}
+
+#endif // PYPY
+
+/** Types with static properties need to handle `Type.static_prop = x` in a specific way.
+    By default, Python replaces the `static_property` itself, but for wrapped C++ types
+    we need to call `static_property.__set__()` in order to propagate the new value to
+    the underlying C++ data structure. */
+extern "C" inline int pybind11_meta_setattro(PyObject *obj, PyObject *name, PyObject *value) {
+    // Use `_PyType_Lookup()` instead of `PyObject_GetAttr()` in order to get the raw
+    // descriptor (`property`) instead of calling `tp_descr_get` (`property.__get__()`).
+    PyObject *descr = _PyType_Lookup((PyTypeObject *) obj, name);
+
+    // The following assignment combinations are possible:
+    //   1. `Type.static_prop = value`             --> descr_set: `Type.static_prop.__set__(value)`
+    //   2. `Type.static_prop = other_static_prop` --> setattro:  replace existing `static_prop`
+    //   3. `Type.regular_attribute = value`       --> setattro:  regular attribute assignment
+    auto *const static_prop = (PyObject *) get_internals().static_property_type;
+    const auto call_descr_set = (descr != nullptr) && (value != nullptr)
+                                && (PyObject_IsInstance(descr, static_prop) != 0)
+                                && (PyObject_IsInstance(value, static_prop) == 0);
+    if (call_descr_set) {
+        // Call `static_property.__set__()` instead of replacing the `static_property`.
+#if !defined(PYPY_VERSION)
+        return Py_TYPE(descr)->tp_descr_set(descr, obj, value);
+#else
+        if (PyObject *result = PyObject_CallMethod(descr, "__set__", "OO", obj, value)) {
+            Py_DECREF(result);
+            return 0;
+        } else {
+            return -1;
+        }
+#endif
+    } else {
+        // Replace existing attribute.
+        return PyType_Type.tp_setattro(obj, name, value);
+    }
+}
+
+/**
+ * Python 3's PyInstanceMethod_Type hides itself via its tp_descr_get, which prevents aliasing
+ * methods via cls.attr("m2") = cls.attr("m1"): instead the tp_descr_get returns a plain function,
+ * when called on a class, or a PyMethod, when called on an instance.  Override that behaviour here
+ * to do a special case bypass for PyInstanceMethod_Types.
+ */
+extern "C" inline PyObject *pybind11_meta_getattro(PyObject *obj, PyObject *name) {
+    PyObject *descr = _PyType_Lookup((PyTypeObject *) obj, name);
+    if (descr && PyInstanceMethod_Check(descr)) {
+        Py_INCREF(descr);
+        return descr;
+    }
+    return PyType_Type.tp_getattro(obj, name);
+}
+
+/// metaclass `__call__` function that is used to create all pybind11 objects.
+extern "C" inline PyObject *pybind11_meta_call(PyObject *type, PyObject *args, PyObject *kwargs) {
+
+    // use the default metaclass call to create/initialize the object
+    PyObject *self = PyType_Type.tp_call(type, args, kwargs);
+    if (self == nullptr) {
+        return nullptr;
+    }
+
+    // Ensure that the base __init__ function(s) were called
+    values_and_holders vhs(self);
+    for (const auto &vh : vhs) {
+        if (!vh.holder_constructed() && !vhs.is_redundant_value_and_holder(vh)) {
+            PyErr_Format(PyExc_TypeError,
+                         "%.200s.__init__() must be called when overriding __init__",
+                         get_fully_qualified_tp_name(vh.type->type).c_str());
+            Py_DECREF(self);
+            return nullptr;
+        }
+    }
+
+    return self;
+}
+
+/// Cleanup the type-info for a pybind11-registered type.
+extern "C" inline void pybind11_meta_dealloc(PyObject *obj) {
+    with_internals([obj](internals &internals) {
+        auto *type = (PyTypeObject *) obj;
+
+        // A pybind11-registered type will:
+        // 1) be found in internals.registered_types_py
+        // 2) have exactly one associated `detail::type_info`
+        auto found_type = internals.registered_types_py.find(type);
+        if (found_type != internals.registered_types_py.end() && found_type->second.size() == 1
+            && found_type->second[0]->type == type) {
+
+            auto *tinfo = found_type->second[0];
+            auto tindex = std::type_index(*tinfo->cpptype);
+            internals.direct_conversions.erase(tindex);
+
+            if (tinfo->module_local) {
+                get_local_internals().registered_types_cpp.erase(tindex);
+            } else {
+                internals.registered_types_cpp.erase(tindex);
+            }
+            internals.registered_types_py.erase(tinfo->type);
+
+            // Actually just `std::erase_if`, but that's only available in C++20
+            auto &cache = internals.inactive_override_cache;
+            for (auto it = cache.begin(), last = cache.end(); it != last;) {
+                if (it->first == (PyObject *) tinfo->type) {
+                    it = cache.erase(it);
+                } else {
+                    ++it;
+                }
+            }
+
+            delete tinfo;
+        }
+    });
+
+    PyType_Type.tp_dealloc(obj);
+}
+
+/** This metaclass is assigned by default to all pybind11 types and is required in order
+    for static properties to function correctly. Users may override this using `py::metaclass`.
+    Return value: New reference. */
+inline PyTypeObject *make_default_metaclass() {
+    constexpr auto *name = "pybind11_type";
+    auto name_obj = reinterpret_steal<object>(PYBIND11_FROM_STRING(name));
+
+    /* Danger zone: from now (and until PyType_Ready), make sure to
+       issue no Python C API calls which could potentially invoke the
+       garbage collector (the GC will call type_traverse(), which will in
+       turn find the newly constructed type in an invalid state) */
+    auto *heap_type = (PyHeapTypeObject *) PyType_Type.tp_alloc(&PyType_Type, 0);
+    if (!heap_type) {
+        pybind11_fail("make_default_metaclass(): error allocating metaclass!");
+    }
+
+    heap_type->ht_name = name_obj.inc_ref().ptr();
+#ifdef PYBIND11_BUILTIN_QUALNAME
+    heap_type->ht_qualname = name_obj.inc_ref().ptr();
+#endif
+
+    auto *type = &heap_type->ht_type;
+    type->tp_name = name;
+    type->tp_base = type_incref(&PyType_Type);
+    type->tp_flags = Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE | Py_TPFLAGS_HEAPTYPE;
+
+    type->tp_call = pybind11_meta_call;
+
+    type->tp_setattro = pybind11_meta_setattro;
+    type->tp_getattro = pybind11_meta_getattro;
+
+    type->tp_dealloc = pybind11_meta_dealloc;
+
+    if (PyType_Ready(type) < 0) {
+        pybind11_fail("make_default_metaclass(): failure in PyType_Ready()!");
+    }
+
+    setattr((PyObject *) type, "__module__", str("pybind11_builtins"));
+    PYBIND11_SET_OLDPY_QUALNAME(type, name_obj);
+
+    return type;
+}
+
+/// For multiple inheritance types we need to recursively register/deregister base pointers for any
+/// base classes with pointers that are difference from the instance value pointer so that we can
+/// correctly recognize an offset base class pointer. This calls a function with any offset base
+/// ptrs.
+inline void traverse_offset_bases(void *valueptr,
+                                  const detail::type_info *tinfo,
+                                  instance *self,
+                                  bool (*f)(void * /*parentptr*/, instance * /*self*/)) {
+    for (handle h : reinterpret_borrow<tuple>(tinfo->type->tp_bases)) {
+        if (auto *parent_tinfo = get_type_info((PyTypeObject *) h.ptr())) {
+            for (auto &c : parent_tinfo->implicit_casts) {
+                if (c.first == tinfo->cpptype) {
+                    auto *parentptr = c.second(valueptr);
+                    if (parentptr != valueptr) {
+                        f(parentptr, self);
+                    }
+                    traverse_offset_bases(parentptr, parent_tinfo, self, f);
+                    break;
+                }
+            }
+        }
+    }
+}
+
+inline bool register_instance_impl(void *ptr, instance *self) {
+    with_instance_map(ptr, [&](instance_map &instances) { instances.emplace(ptr, self); });
+    return true; // unused, but gives the same signature as the deregister func
+}
+inline bool deregister_instance_impl(void *ptr, instance *self) {
+    return with_instance_map(ptr, [&](instance_map &instances) {
+        auto range = instances.equal_range(ptr);
+        for (auto it = range.first; it != range.second; ++it) {
+            if (self == it->second) {
+                instances.erase(it);
+                return true;
+            }
+        }
+        return false;
+    });
+}
+
+inline void register_instance(instance *self, void *valptr, const type_info *tinfo) {
+    register_instance_impl(valptr, self);
+    if (!tinfo->simple_ancestors) {
+        traverse_offset_bases(valptr, tinfo, self, register_instance_impl);
+    }
+}
+
+inline bool deregister_instance(instance *self, void *valptr, const type_info *tinfo) {
+    bool ret = deregister_instance_impl(valptr, self);
+    if (!tinfo->simple_ancestors) {
+        traverse_offset_bases(valptr, tinfo, self, deregister_instance_impl);
+    }
+    return ret;
+}
+
+/// Instance creation function for all pybind11 types. It allocates the internal instance layout
+/// for holding C++ objects and holders.  Allocation is done lazily (the first time the instance is
+/// cast to a reference or pointer), and initialization is done by an `__init__` function.
+inline PyObject *make_new_instance(PyTypeObject *type) {
+#if defined(PYPY_VERSION)
+    // PyPy gets tp_basicsize wrong (issue 2482) under multiple inheritance when the first
+    // inherited object is a plain Python type (i.e. not derived from an extension type).  Fix it.
+    ssize_t instance_size = static_cast<ssize_t>(sizeof(instance));
+    if (type->tp_basicsize < instance_size) {
+        type->tp_basicsize = instance_size;
+    }
+#endif
+    PyObject *self = type->tp_alloc(type, 0);
+    auto *inst = reinterpret_cast<instance *>(self);
+    // Allocate the value/holder internals:
+    inst->allocate_layout();
+
+    return self;
+}
+
+/// Instance creation function for all pybind11 types. It only allocates space for the
+/// C++ object, but doesn't call the constructor -- an `__init__` function must do that.
+extern "C" inline PyObject *pybind11_object_new(PyTypeObject *type, PyObject *, PyObject *) {
+    return make_new_instance(type);
+}
+
+/// An `__init__` function constructs the C++ object. Users should provide at least one
+/// of these using `py::init` or directly with `.def(__init__, ...)`. Otherwise, the
+/// following default function will be used which simply throws an exception.
+extern "C" inline int pybind11_object_init(PyObject *self, PyObject *, PyObject *) {
+    PyTypeObject *type = Py_TYPE(self);
+    std::string msg = get_fully_qualified_tp_name(type) + ": No constructor defined!";
+    set_error(PyExc_TypeError, msg.c_str());
+    return -1;
+}
+
+inline void add_patient(PyObject *nurse, PyObject *patient) {
+    auto *instance = reinterpret_cast<detail::instance *>(nurse);
+    instance->has_patients = true;
+    Py_INCREF(patient);
+
+    with_internals([&](internals &internals) { internals.patients[nurse].push_back(patient); });
+}
+
+inline void clear_patients(PyObject *self) {
+    auto *instance = reinterpret_cast<detail::instance *>(self);
+    std::vector<PyObject *> patients;
+
+    with_internals([&](internals &internals) {
+        auto pos = internals.patients.find(self);
+
+        if (pos == internals.patients.end()) {
+            pybind11_fail(
+                "FATAL: Internal consistency check failed: Invalid clear_patients() call.");
+        }
+
+        // Clearing the patients can cause more Python code to run, which
+        // can invalidate the iterator. Extract the vector of patients
+        // from the unordered_map first.
+        patients = std::move(pos->second);
+        internals.patients.erase(pos);
+    });
+
+    instance->has_patients = false;
+    for (PyObject *&patient : patients) {
+        Py_CLEAR(patient);
+    }
+}
+
+/// Clears all internal data from the instance and removes it from registered instances in
+/// preparation for deallocation.
+inline void clear_instance(PyObject *self) {
+    auto *instance = reinterpret_cast<detail::instance *>(self);
+
+    // Deallocate any values/holders, if present:
+    for (auto &v_h : values_and_holders(instance)) {
+        if (v_h) {
+
+            // We have to deregister before we call dealloc because, for virtual MI types, we still
+            // need to be able to get the parent pointers.
+            if (v_h.instance_registered()
+                && !deregister_instance(instance, v_h.value_ptr(), v_h.type)) {
+                pybind11_fail(
+                    "pybind11_object_dealloc(): Tried to deallocate unregistered instance!");
+            }
+
+            if (instance->owned || v_h.holder_constructed()) {
+                v_h.type->dealloc(v_h);
+            }
+        }
+    }
+    // Deallocate the value/holder layout internals:
+    instance->deallocate_layout();
+
+    if (instance->weakrefs) {
+        PyObject_ClearWeakRefs(self);
+    }
+
+    PyObject **dict_ptr = _PyObject_GetDictPtr(self);
+    if (dict_ptr) {
+        Py_CLEAR(*dict_ptr);
+    }
+
+    if (instance->has_patients) {
+        clear_patients(self);
+    }
+}
+
+/// Instance destructor function for all pybind11 types. It calls `type_info.dealloc`
+/// to destroy the C++ object itself, while the rest is Python bookkeeping.
+extern "C" inline void pybind11_object_dealloc(PyObject *self) {
+    auto *type = Py_TYPE(self);
+
+    // If this is a GC tracked object, untrack it first
+    // Note that the track call is implicitly done by the
+    // default tp_alloc, which we never override.
+    if (PyType_HasFeature(type, Py_TPFLAGS_HAVE_GC) != 0) {
+        PyObject_GC_UnTrack(self);
+    }
+
+    clear_instance(self);
+
+    type->tp_free(self);
+
+#if PY_VERSION_HEX < 0x03080000
+    // `type->tp_dealloc != pybind11_object_dealloc` means that we're being called
+    // as part of a derived type's dealloc, in which case we're not allowed to decref
+    // the type here. For cross-module compatibility, we shouldn't compare directly
+    // with `pybind11_object_dealloc`, but with the common one stashed in internals.
+    auto pybind11_object_type = (PyTypeObject *) get_internals().instance_base;
+    if (type->tp_dealloc == pybind11_object_type->tp_dealloc)
+        Py_DECREF(type);
+#else
+    // This was not needed before Python 3.8 (Python issue 35810)
+    // https://github.com/pybind/pybind11/issues/1946
+    Py_DECREF(type);
+#endif
+}
+
+std::string error_string();
+
+/** Create the type which can be used as a common base for all classes.  This is
+    needed in order to satisfy Python's requirements for multiple inheritance.
+    Return value: New reference. */
+inline PyObject *make_object_base_type(PyTypeObject *metaclass) {
+    constexpr auto *name = "pybind11_object";
+    auto name_obj = reinterpret_steal<object>(PYBIND11_FROM_STRING(name));
+
+    /* Danger zone: from now (and until PyType_Ready), make sure to
+       issue no Python C API calls which could potentially invoke the
+       garbage collector (the GC will call type_traverse(), which will in
+       turn find the newly constructed type in an invalid state) */
+    auto *heap_type = (PyHeapTypeObject *) metaclass->tp_alloc(metaclass, 0);
+    if (!heap_type) {
+        pybind11_fail("make_object_base_type(): error allocating type!");
+    }
+
+    heap_type->ht_name = name_obj.inc_ref().ptr();
+#ifdef PYBIND11_BUILTIN_QUALNAME
+    heap_type->ht_qualname = name_obj.inc_ref().ptr();
+#endif
+
+    auto *type = &heap_type->ht_type;
+    type->tp_name = name;
+    type->tp_base = type_incref(&PyBaseObject_Type);
+    type->tp_basicsize = static_cast<ssize_t>(sizeof(instance));
+    type->tp_flags = Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE | Py_TPFLAGS_HEAPTYPE;
+
+    type->tp_new = pybind11_object_new;
+    type->tp_init = pybind11_object_init;
+    type->tp_dealloc = pybind11_object_dealloc;
+
+    /* Support weak references (needed for the keep_alive feature) */
+    type->tp_weaklistoffset = offsetof(instance, weakrefs);
+
+    if (PyType_Ready(type) < 0) {
+        pybind11_fail("PyType_Ready failed in make_object_base_type(): " + error_string());
+    }
+
+    setattr((PyObject *) type, "__module__", str("pybind11_builtins"));
+    PYBIND11_SET_OLDPY_QUALNAME(type, name_obj);
+
+    assert(!PyType_HasFeature(type, Py_TPFLAGS_HAVE_GC));
+    return (PyObject *) heap_type;
+}
+
+/// dynamic_attr: Allow the garbage collector to traverse the internal instance `__dict__`.
+extern "C" inline int pybind11_traverse(PyObject *self, visitproc visit, void *arg) {
+#if PY_VERSION_HEX >= 0x030D0000
+    PyObject_VisitManagedDict(self, visit, arg);
+#else
+    PyObject *&dict = *_PyObject_GetDictPtr(self);
+    Py_VISIT(dict);
+#endif
+// https://docs.python.org/3/c-api/typeobj.html#c.PyTypeObject.tp_traverse
+#if PY_VERSION_HEX >= 0x03090000
+    Py_VISIT(Py_TYPE(self));
+#endif
+    return 0;
+}
+
+/// dynamic_attr: Allow the GC to clear the dictionary.
+extern "C" inline int pybind11_clear(PyObject *self) {
+#if PY_VERSION_HEX >= 0x030D0000
+    PyObject_ClearManagedDict(self);
+#else
+    PyObject *&dict = *_PyObject_GetDictPtr(self);
+    Py_CLEAR(dict);
+#endif
+    return 0;
+}
+
+/// Give instances of this type a `__dict__` and opt into garbage collection.
+inline void enable_dynamic_attributes(PyHeapTypeObject *heap_type) {
+    auto *type = &heap_type->ht_type;
+    type->tp_flags |= Py_TPFLAGS_HAVE_GC;
+#if PY_VERSION_HEX < 0x030B0000
+    type->tp_dictoffset = type->tp_basicsize;           // place dict at the end
+    type->tp_basicsize += (ssize_t) sizeof(PyObject *); // and allocate enough space for it
+#else
+    type->tp_flags |= Py_TPFLAGS_MANAGED_DICT;
+#endif
+    type->tp_traverse = pybind11_traverse;
+    type->tp_clear = pybind11_clear;
+
+    static PyGetSetDef getset[]
+        = {{"__dict__", PyObject_GenericGetDict, PyObject_GenericSetDict, nullptr, nullptr},
+           {nullptr, nullptr, nullptr, nullptr, nullptr}};
+    type->tp_getset = getset;
+}
+
+/// buffer_protocol: Fill in the view as specified by flags.
+extern "C" inline int pybind11_getbuffer(PyObject *obj, Py_buffer *view, int flags) {
+    // Look for a `get_buffer` implementation in this type's info or any bases (following MRO).
+    type_info *tinfo = nullptr;
+    for (auto type : reinterpret_borrow<tuple>(Py_TYPE(obj)->tp_mro)) {
+        tinfo = get_type_info((PyTypeObject *) type.ptr());
+        if (tinfo && tinfo->get_buffer) {
+            break;
+        }
+    }
+    if (view == nullptr || !tinfo || !tinfo->get_buffer) {
+        if (view) {
+            view->obj = nullptr;
+        }
+        set_error(PyExc_BufferError, "pybind11_getbuffer(): Internal error");
+        return -1;
+    }
+    std::memset(view, 0, sizeof(Py_buffer));
+    buffer_info *info = nullptr;
+    try {
+        info = tinfo->get_buffer(obj, tinfo->get_buffer_data);
+    } catch (...) {
+        try_translate_exceptions();
+        raise_from(PyExc_BufferError, "Error getting buffer");
+        return -1;
+    }
+    if (info == nullptr) {
+        pybind11_fail("FATAL UNEXPECTED SITUATION: tinfo->get_buffer() returned nullptr.");
+    }
+
+    if ((flags & PyBUF_WRITABLE) == PyBUF_WRITABLE && info->readonly) {
+        delete info;
+        // view->obj = nullptr;  // Was just memset to 0, so not necessary
+        set_error(PyExc_BufferError, "Writable buffer requested for readonly storage");
+        return -1;
+    }
+    view->obj = obj;
+    view->ndim = 1;
+    view->internal = info;
+    view->buf = info->ptr;
+    view->itemsize = info->itemsize;
+    view->len = view->itemsize;
+    for (auto s : info->shape) {
+        view->len *= s;
+    }
+    view->readonly = static_cast<int>(info->readonly);
+    if ((flags & PyBUF_FORMAT) == PyBUF_FORMAT) {
+        view->format = const_cast<char *>(info->format.c_str());
+    }
+    if ((flags & PyBUF_STRIDES) == PyBUF_STRIDES) {
+        view->ndim = (int) info->ndim;
+        view->strides = info->strides.data();
+        view->shape = info->shape.data();
+    }
+    Py_INCREF(view->obj);
+    return 0;
+}
+
+/// buffer_protocol: Release the resources of the buffer.
+extern "C" inline void pybind11_releasebuffer(PyObject *, Py_buffer *view) {
+    delete (buffer_info *) view->internal;
+}
+
+/// Give this type a buffer interface.
+inline void enable_buffer_protocol(PyHeapTypeObject *heap_type) {
+    heap_type->ht_type.tp_as_buffer = &heap_type->as_buffer;
+
+    heap_type->as_buffer.bf_getbuffer = pybind11_getbuffer;
+    heap_type->as_buffer.bf_releasebuffer = pybind11_releasebuffer;
+}
+
+/** Create a brand new Python type according to the `type_record` specification.
+    Return value: New reference. */
+inline PyObject *make_new_python_type(const type_record &rec) {
+    auto name = reinterpret_steal<object>(PYBIND11_FROM_STRING(rec.name));
+
+    auto qualname = name;
+    if (rec.scope && !PyModule_Check(rec.scope.ptr()) && hasattr(rec.scope, "__qualname__")) {
+        qualname = reinterpret_steal<object>(
+            PyUnicode_FromFormat("%U.%U", rec.scope.attr("__qualname__").ptr(), name.ptr()));
+    }
+
+    object module_;
+    if (rec.scope) {
+        if (hasattr(rec.scope, "__module__")) {
+            module_ = rec.scope.attr("__module__");
+        } else if (hasattr(rec.scope, "__name__")) {
+            module_ = rec.scope.attr("__name__");
+        }
+    }
+
+    const auto *full_name = c_str(
+#if !defined(PYPY_VERSION)
+        module_ ? str(module_).cast<std::string>() + "." + rec.name :
+#endif
+                rec.name);
+
+    char *tp_doc = nullptr;
+    if (rec.doc && options::show_user_defined_docstrings()) {
+        /* Allocate memory for docstring (Python will free this later on) */
+        size_t size = std::strlen(rec.doc) + 1;
+#if PY_VERSION_HEX >= 0x030D0000
+        tp_doc = (char *) PyMem_MALLOC(size);
+#else
+        tp_doc = (char *) PyObject_MALLOC(size);
+#endif
+        std::memcpy((void *) tp_doc, rec.doc, size);
+    }
+
+    auto &internals = get_internals();
+    auto bases = tuple(rec.bases);
+    auto *base = (bases.empty()) ? internals.instance_base : bases[0].ptr();
+
+    /* Danger zone: from now (and until PyType_Ready), make sure to
+       issue no Python C API calls which could potentially invoke the
+       garbage collector (the GC will call type_traverse(), which will in
+       turn find the newly constructed type in an invalid state) */
+    auto *metaclass
+        = rec.metaclass.ptr() ? (PyTypeObject *) rec.metaclass.ptr() : internals.default_metaclass;
+
+    auto *heap_type = (PyHeapTypeObject *) metaclass->tp_alloc(metaclass, 0);
+    if (!heap_type) {
+        pybind11_fail(std::string(rec.name) + ": Unable to create type object!");
+    }
+
+    heap_type->ht_name = name.release().ptr();
+#ifdef PYBIND11_BUILTIN_QUALNAME
+    heap_type->ht_qualname = qualname.inc_ref().ptr();
+#endif
+
+    auto *type = &heap_type->ht_type;
+    type->tp_name = full_name;
+    type->tp_doc = tp_doc;
+    type->tp_base = type_incref((PyTypeObject *) base);
+    type->tp_basicsize = static_cast<ssize_t>(sizeof(instance));
+    if (!bases.empty()) {
+        type->tp_bases = bases.release().ptr();
+    }
+
+    /* Don't inherit base __init__ */
+    type->tp_init = pybind11_object_init;
+
+    /* Supported protocols */
+    type->tp_as_number = &heap_type->as_number;
+    type->tp_as_sequence = &heap_type->as_sequence;
+    type->tp_as_mapping = &heap_type->as_mapping;
+    type->tp_as_async = &heap_type->as_async;
+
+    /* Flags */
+    type->tp_flags |= Py_TPFLAGS_DEFAULT | Py_TPFLAGS_HEAPTYPE;
+    if (!rec.is_final) {
+        type->tp_flags |= Py_TPFLAGS_BASETYPE;
+    }
+
+    if (rec.dynamic_attr) {
+        enable_dynamic_attributes(heap_type);
+    }
+
+    if (rec.buffer_protocol) {
+        enable_buffer_protocol(heap_type);
+    }
+
+    if (rec.custom_type_setup_callback) {
+        rec.custom_type_setup_callback(heap_type);
+    }
+
+    if (PyType_Ready(type) < 0) {
+        pybind11_fail(std::string(rec.name) + ": PyType_Ready failed: " + error_string());
+    }
+
+    assert(!rec.dynamic_attr || PyType_HasFeature(type, Py_TPFLAGS_HAVE_GC));
+
+    /* Register type with the parent scope */
+    if (rec.scope) {
+        setattr(rec.scope, rec.name, (PyObject *) type);
+    } else {
+        Py_INCREF(type); // Keep it alive forever (reference leak)
+    }
+
+    if (module_) { // Needed by pydoc
+        setattr((PyObject *) type, "__module__", module_);
+    }
+
+    PYBIND11_SET_OLDPY_QUALNAME(type, qualname);
+
+    return (PyObject *) type;
+}
+
+PYBIND11_NAMESPACE_END(detail)
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/detail/common.h

@@ -0,0 +1,1287 @@
+/*
+    pybind11/detail/common.h -- Basic macros
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#define PYBIND11_VERSION_MAJOR 2
+#define PYBIND11_VERSION_MINOR 13
+#define PYBIND11_VERSION_PATCH 6
+
+// Similar to Python's convention: https://docs.python.org/3/c-api/apiabiversion.html
+// Additional convention: 0xD = dev
+#define PYBIND11_VERSION_HEX 0x020D0600
+
+// Define some generic pybind11 helper macros for warning management.
+//
+// Note that compiler-specific push/pop pairs are baked into the
+// PYBIND11_NAMESPACE_BEGIN/PYBIND11_NAMESPACE_END pair of macros. Therefore manual
+// PYBIND11_WARNING_PUSH/PYBIND11_WARNING_POP are usually only needed in `#include` sections.
+//
+// If you find you need to suppress a warning, please try to make the suppression as local as
+// possible using these macros. Please also be sure to push/pop with the pybind11 macros. Please
+// only use compiler specifics if you need to check specific versions, e.g. Apple Clang vs. vanilla
+// Clang.
+#if defined(_MSC_VER)
+#    define PYBIND11_COMPILER_MSVC
+#    define PYBIND11_PRAGMA(...) __pragma(__VA_ARGS__)
+#    define PYBIND11_WARNING_PUSH PYBIND11_PRAGMA(warning(push))
+#    define PYBIND11_WARNING_POP PYBIND11_PRAGMA(warning(pop))
+#elif defined(__INTEL_COMPILER)
+#    define PYBIND11_COMPILER_INTEL
+#    define PYBIND11_PRAGMA(...) _Pragma(#__VA_ARGS__)
+#    define PYBIND11_WARNING_PUSH PYBIND11_PRAGMA(warning push)
+#    define PYBIND11_WARNING_POP PYBIND11_PRAGMA(warning pop)
+#elif defined(__clang__)
+#    define PYBIND11_COMPILER_CLANG
+#    define PYBIND11_PRAGMA(...) _Pragma(#__VA_ARGS__)
+#    define PYBIND11_WARNING_PUSH PYBIND11_PRAGMA(clang diagnostic push)
+#    define PYBIND11_WARNING_POP PYBIND11_PRAGMA(clang diagnostic push)
+#elif defined(__GNUC__)
+#    define PYBIND11_COMPILER_GCC
+#    define PYBIND11_PRAGMA(...) _Pragma(#__VA_ARGS__)
+#    define PYBIND11_WARNING_PUSH PYBIND11_PRAGMA(GCC diagnostic push)
+#    define PYBIND11_WARNING_POP PYBIND11_PRAGMA(GCC diagnostic pop)
+#endif
+
+#ifdef PYBIND11_COMPILER_MSVC
+#    define PYBIND11_WARNING_DISABLE_MSVC(name) PYBIND11_PRAGMA(warning(disable : name))
+#else
+#    define PYBIND11_WARNING_DISABLE_MSVC(name)
+#endif
+
+#ifdef PYBIND11_COMPILER_CLANG
+#    define PYBIND11_WARNING_DISABLE_CLANG(name) PYBIND11_PRAGMA(clang diagnostic ignored name)
+#else
+#    define PYBIND11_WARNING_DISABLE_CLANG(name)
+#endif
+
+#ifdef PYBIND11_COMPILER_GCC
+#    define PYBIND11_WARNING_DISABLE_GCC(name) PYBIND11_PRAGMA(GCC diagnostic ignored name)
+#else
+#    define PYBIND11_WARNING_DISABLE_GCC(name)
+#endif
+
+#ifdef PYBIND11_COMPILER_INTEL
+#    define PYBIND11_WARNING_DISABLE_INTEL(name) PYBIND11_PRAGMA(warning disable name)
+#else
+#    define PYBIND11_WARNING_DISABLE_INTEL(name)
+#endif
+
+#define PYBIND11_NAMESPACE_BEGIN(name)                                                            \
+    namespace name {                                                                              \
+    PYBIND11_WARNING_PUSH
+
+#define PYBIND11_NAMESPACE_END(name)                                                              \
+    PYBIND11_WARNING_POP                                                                          \
+    }
+
+// Robust support for some features and loading modules compiled against different pybind versions
+// requires forcing hidden visibility on pybind code, so we enforce this by setting the attribute
+// on the main `pybind11` namespace.
+#if !defined(PYBIND11_NAMESPACE)
+#    ifdef __GNUG__
+#        define PYBIND11_NAMESPACE pybind11 __attribute__((visibility("hidden")))
+#    else
+#        define PYBIND11_NAMESPACE pybind11
+#    endif
+#endif
+
+#if !(defined(_MSC_VER) && __cplusplus == 199711L)
+#    if __cplusplus >= 201402L
+#        define PYBIND11_CPP14
+#        if __cplusplus >= 201703L
+#            define PYBIND11_CPP17
+#            if __cplusplus >= 202002L
+#                define PYBIND11_CPP20
+// Please update tests/pybind11_tests.cpp `cpp_std()` when adding a macro here.
+#            endif
+#        endif
+#    endif
+#elif defined(_MSC_VER) && __cplusplus == 199711L
+// MSVC sets _MSVC_LANG rather than __cplusplus (supposedly until the standard is fully
+// implemented). Unless you use the /Zc:__cplusplus flag on Visual Studio 2017 15.7 Preview 3
+// or newer.
+#    if _MSVC_LANG >= 201402L
+#        define PYBIND11_CPP14
+#        if _MSVC_LANG > 201402L
+#            define PYBIND11_CPP17
+#            if _MSVC_LANG >= 202002L
+#                define PYBIND11_CPP20
+#            endif
+#        endif
+#    endif
+#endif
+
+#if defined(PYBIND11_CPP20)
+#    define PYBIND11_CONSTINIT constinit
+#    define PYBIND11_DTOR_CONSTEXPR constexpr
+#else
+#    define PYBIND11_CONSTINIT
+#    define PYBIND11_DTOR_CONSTEXPR
+#endif
+
+// Compiler version assertions
+#if defined(__INTEL_COMPILER)
+#    if __INTEL_COMPILER < 1800
+#        error pybind11 requires Intel C++ compiler v18 or newer
+#    elif __INTEL_COMPILER < 1900 && defined(PYBIND11_CPP14)
+#        error pybind11 supports only C++11 with Intel C++ compiler v18. Use v19 or newer for C++14.
+#    endif
+/* The following pragma cannot be pop'ed:
+   https://community.intel.com/t5/Intel-C-Compiler/Inline-and-no-inline-warning/td-p/1216764 */
+#    pragma warning disable 2196 // warning #2196: routine is both "inline" and "noinline"
+#elif defined(__clang__) && !defined(__apple_build_version__)
+#    if __clang_major__ < 3 || (__clang_major__ == 3 && __clang_minor__ < 3)
+#        error pybind11 requires clang 3.3 or newer
+#    endif
+#elif defined(__clang__)
+// Apple changes clang version macros to its Xcode version; the first Xcode release based on
+// (upstream) clang 3.3 was Xcode 5:
+#    if __clang_major__ < 5
+#        error pybind11 requires Xcode/clang 5.0 or newer
+#    endif
+#elif defined(__GNUG__)
+#    if __GNUC__ < 4 || (__GNUC__ == 4 && __GNUC_MINOR__ < 8)
+#        error pybind11 requires gcc 4.8 or newer
+#    endif
+#elif defined(_MSC_VER)
+#    if _MSC_VER < 1910
+#        error pybind11 2.10+ requires MSVC 2017 or newer
+#    endif
+#endif
+
+#if !defined(PYBIND11_EXPORT)
+#    if defined(WIN32) || defined(_WIN32)
+#        define PYBIND11_EXPORT __declspec(dllexport)
+#    else
+#        define PYBIND11_EXPORT __attribute__((visibility("default")))
+#    endif
+#endif
+
+#if !defined(PYBIND11_EXPORT_EXCEPTION)
+#    if defined(__apple_build_version__)
+#        define PYBIND11_EXPORT_EXCEPTION PYBIND11_EXPORT
+#    else
+#        define PYBIND11_EXPORT_EXCEPTION
+#    endif
+#endif
+
+// For CUDA, GCC7, GCC8:
+// PYBIND11_NOINLINE_FORCED is incompatible with `-Wattributes -Werror`.
+// When defining PYBIND11_NOINLINE_FORCED, it is best to also use `-Wno-attributes`.
+// However, the measured shared-library size saving when using noinline are only
+// 1.7% for CUDA, -0.2% for GCC7, and 0.0% for GCC8 (using -DCMAKE_BUILD_TYPE=MinSizeRel,
+// the default under pybind11/tests).
+#if !defined(PYBIND11_NOINLINE_FORCED)                                                            \
+    && (defined(__CUDACC__) || (defined(__GNUC__) && (__GNUC__ == 7 || __GNUC__ == 8)))
+#    define PYBIND11_NOINLINE_DISABLED
+#endif
+
+// The PYBIND11_NOINLINE macro is for function DEFINITIONS.
+// In contrast, FORWARD DECLARATIONS should never use this macro:
+// https://stackoverflow.com/questions/9317473/forward-declaration-of-inline-functions
+#if defined(PYBIND11_NOINLINE_DISABLED) // Option for maximum portability and experimentation.
+#    define PYBIND11_NOINLINE inline
+#elif defined(_MSC_VER)
+#    define PYBIND11_NOINLINE __declspec(noinline) inline
+#else
+#    define PYBIND11_NOINLINE __attribute__((noinline)) inline
+#endif
+
+#if defined(__MINGW32__)
+// For unknown reasons all PYBIND11_DEPRECATED member trigger a warning when declared
+// whether it is used or not
+#    define PYBIND11_DEPRECATED(reason)
+#elif defined(PYBIND11_CPP14)
+#    define PYBIND11_DEPRECATED(reason) [[deprecated(reason)]]
+#else
+#    define PYBIND11_DEPRECATED(reason) __attribute__((deprecated(reason)))
+#endif
+
+#if defined(PYBIND11_CPP17)
+#    define PYBIND11_MAYBE_UNUSED [[maybe_unused]]
+#elif defined(_MSC_VER) && !defined(__clang__)
+#    define PYBIND11_MAYBE_UNUSED
+#else
+#    define PYBIND11_MAYBE_UNUSED __attribute__((__unused__))
+#endif
+
+/* Don't let Python.h #define (v)snprintf as macro because they are implemented
+   properly in Visual Studio since 2015. */
+#if defined(_MSC_VER)
+#    define HAVE_SNPRINTF 1
+#endif
+
+/// Include Python header, disable linking to pythonX_d.lib on Windows in debug mode
+#if defined(_MSC_VER)
+PYBIND11_WARNING_PUSH
+PYBIND11_WARNING_DISABLE_MSVC(4505)
+// C4505: 'PySlice_GetIndicesEx': unreferenced local function has been removed (PyPy only)
+#    if defined(_DEBUG) && !defined(Py_DEBUG)
+// Workaround for a VS 2022 issue.
+// NOTE: This workaround knowingly violates the Python.h include order requirement:
+// https://docs.python.org/3/c-api/intro.html#include-files
+// See https://github.com/pybind/pybind11/pull/3497 for full context.
+#        include <yvals.h>
+#        if _MSVC_STL_VERSION >= 143
+#            include <crtdefs.h>
+#        endif
+#        define PYBIND11_DEBUG_MARKER
+#        undef _DEBUG
+#    endif
+#endif
+
+// https://en.cppreference.com/w/c/chrono/localtime
+#if defined(__STDC_LIB_EXT1__) && !defined(__STDC_WANT_LIB_EXT1__)
+#    define __STDC_WANT_LIB_EXT1__
+#endif
+
+#ifdef __has_include
+// std::optional (but including it in c++14 mode isn't allowed)
+#    if defined(PYBIND11_CPP17) && __has_include(<optional>)
+#        define PYBIND11_HAS_OPTIONAL 1
+#    endif
+// std::experimental::optional (but not allowed in c++11 mode)
+#    if defined(PYBIND11_CPP14) && (__has_include(<experimental/optional>) && \
+                                 !__has_include(<optional>))
+#        define PYBIND11_HAS_EXP_OPTIONAL 1
+#    endif
+// std::variant
+#    if defined(PYBIND11_CPP17) && __has_include(<variant>)
+#        define PYBIND11_HAS_VARIANT 1
+#    endif
+#elif defined(_MSC_VER) && defined(PYBIND11_CPP17)
+#    define PYBIND11_HAS_OPTIONAL 1
+#    define PYBIND11_HAS_VARIANT 1
+#endif
+
+#if defined(PYBIND11_CPP17)
+#    if defined(__has_include)
+#        if __has_include(<string_view>)
+#            define PYBIND11_HAS_STRING_VIEW
+#        endif
+#    elif defined(_MSC_VER)
+#        define PYBIND11_HAS_STRING_VIEW
+#    endif
+#endif
+
+#include <Python.h>
+#if PY_VERSION_HEX < 0x03070000
+#    error "PYTHON < 3.7 IS UNSUPPORTED. pybind11 v2.12 was the last to support Python 3.6."
+#endif
+#include <frameobject.h>
+#include <pythread.h>
+
+/* Python #defines overrides on all sorts of core functions, which
+   tends to weak havok in C++ codebases that expect these to work
+   like regular functions (potentially with several overloads) */
+#if defined(isalnum)
+#    undef isalnum
+#    undef isalpha
+#    undef islower
+#    undef isspace
+#    undef isupper
+#    undef tolower
+#    undef toupper
+#endif
+
+#if defined(copysign)
+#    undef copysign
+#endif
+
+#if defined(PYBIND11_NUMPY_1_ONLY)
+#    define PYBIND11_INTERNAL_NUMPY_1_ONLY_DETECTED
+#endif
+
+#if defined(PYPY_VERSION) && !defined(PYBIND11_SIMPLE_GIL_MANAGEMENT)
+#    define PYBIND11_SIMPLE_GIL_MANAGEMENT
+#endif
+
+#if defined(_MSC_VER)
+#    if defined(PYBIND11_DEBUG_MARKER)
+#        define _DEBUG
+#        undef PYBIND11_DEBUG_MARKER
+#    endif
+PYBIND11_WARNING_POP
+#endif
+
+#include <cstddef>
+#include <cstring>
+#include <exception>
+#include <forward_list>
+#include <memory>
+#include <stdexcept>
+#include <string>
+#include <type_traits>
+#include <typeindex>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+#if defined(__has_include)
+#    if __has_include(<version>)
+#        include <version>
+#    endif
+#endif
+
+// Must be after including <version> or one of the other headers specified by the standard
+#if defined(__cpp_lib_char8_t) && __cpp_lib_char8_t >= 201811L
+#    define PYBIND11_HAS_U8STRING
+#endif
+
+// See description of PR #4246:
+#if !defined(PYBIND11_NO_ASSERT_GIL_HELD_INCREF_DECREF) && !defined(NDEBUG)                       \
+    && !defined(PYPY_VERSION) && !defined(PYBIND11_ASSERT_GIL_HELD_INCREF_DECREF)
+#    define PYBIND11_ASSERT_GIL_HELD_INCREF_DECREF
+#endif
+
+// #define PYBIND11_STR_LEGACY_PERMISSIVE
+// If DEFINED, pybind11::str can hold PyUnicodeObject or PyBytesObject
+//             (probably surprising and never documented, but this was the
+//             legacy behavior until and including v2.6.x). As a side-effect,
+//             pybind11::isinstance<str>() is true for both pybind11::str and
+//             pybind11::bytes.
+// If UNDEFINED, pybind11::str can only hold PyUnicodeObject, and
+//               pybind11::isinstance<str>() is true only for pybind11::str.
+//               However, for Python 2 only (!), the pybind11::str caster
+//               implicitly decoded bytes to PyUnicodeObject. This was to ease
+//               the transition from the legacy behavior to the non-permissive
+//               behavior.
+
+/// Compatibility macros for Python 2 / Python 3 versions TODO: remove
+#define PYBIND11_INSTANCE_METHOD_NEW(ptr, class_) PyInstanceMethod_New(ptr)
+#define PYBIND11_INSTANCE_METHOD_CHECK PyInstanceMethod_Check
+#define PYBIND11_INSTANCE_METHOD_GET_FUNCTION PyInstanceMethod_GET_FUNCTION
+#define PYBIND11_BYTES_CHECK PyBytes_Check
+#define PYBIND11_BYTES_FROM_STRING PyBytes_FromString
+#define PYBIND11_BYTES_FROM_STRING_AND_SIZE PyBytes_FromStringAndSize
+#define PYBIND11_BYTES_AS_STRING_AND_SIZE PyBytes_AsStringAndSize
+#define PYBIND11_BYTES_AS_STRING PyBytes_AsString
+#define PYBIND11_BYTES_SIZE PyBytes_Size
+#define PYBIND11_LONG_CHECK(o) PyLong_Check(o)
+#define PYBIND11_LONG_AS_LONGLONG(o) PyLong_AsLongLong(o)
+#define PYBIND11_LONG_FROM_SIGNED(o) PyLong_FromSsize_t((ssize_t) (o))
+#define PYBIND11_LONG_FROM_UNSIGNED(o) PyLong_FromSize_t((size_t) (o))
+#define PYBIND11_BYTES_NAME "bytes"
+#define PYBIND11_STRING_NAME "str"
+#define PYBIND11_SLICE_OBJECT PyObject
+#define PYBIND11_FROM_STRING PyUnicode_FromString
+#define PYBIND11_STR_TYPE ::pybind11::str
+#define PYBIND11_BOOL_ATTR "__bool__"
+#define PYBIND11_NB_BOOL(ptr) ((ptr)->nb_bool)
+#define PYBIND11_BUILTINS_MODULE "builtins"
+// Providing a separate declaration to make Clang's -Wmissing-prototypes happy.
+// See comment for PYBIND11_MODULE below for why this is marked "maybe unused".
+#define PYBIND11_PLUGIN_IMPL(name)                                                                \
+    extern "C" PYBIND11_MAYBE_UNUSED PYBIND11_EXPORT PyObject *PyInit_##name();                   \
+    extern "C" PYBIND11_EXPORT PyObject *PyInit_##name()
+
+#define PYBIND11_TRY_NEXT_OVERLOAD ((PyObject *) 1) // special failure return code
+#define PYBIND11_STRINGIFY(x) #x
+#define PYBIND11_TOSTRING(x) PYBIND11_STRINGIFY(x)
+#define PYBIND11_CONCAT(first, second) first##second
+#define PYBIND11_ENSURE_INTERNALS_READY pybind11::detail::get_internals();
+
+#define PYBIND11_CHECK_PYTHON_VERSION                                                             \
+    {                                                                                             \
+        const char *compiled_ver                                                                  \
+            = PYBIND11_TOSTRING(PY_MAJOR_VERSION) "." PYBIND11_TOSTRING(PY_MINOR_VERSION);        \
+        const char *runtime_ver = Py_GetVersion();                                                \
+        size_t len = std::strlen(compiled_ver);                                                   \
+        if (std::strncmp(runtime_ver, compiled_ver, len) != 0                                     \
+            || (runtime_ver[len] >= '0' && runtime_ver[len] <= '9')) {                            \
+            PyErr_Format(PyExc_ImportError,                                                       \
+                         "Python version mismatch: module was compiled for Python %s, "           \
+                         "but the interpreter version is incompatible: %s.",                      \
+                         compiled_ver,                                                            \
+                         runtime_ver);                                                            \
+            return nullptr;                                                                       \
+        }                                                                                         \
+    }
+
+#define PYBIND11_CATCH_INIT_EXCEPTIONS                                                            \
+    catch (pybind11::error_already_set & e) {                                                     \
+        pybind11::raise_from(e, PyExc_ImportError, "initialization failed");                      \
+        return nullptr;                                                                           \
+    }                                                                                             \
+    catch (const std::exception &e) {                                                             \
+        ::pybind11::set_error(PyExc_ImportError, e.what());                                       \
+        return nullptr;                                                                           \
+    }
+
+/** \rst
+    ***Deprecated in favor of PYBIND11_MODULE***
+
+    This macro creates the entry point that will be invoked when the Python interpreter
+    imports a plugin library. Please create a `module_` in the function body and return
+    the pointer to its underlying Python object at the end.
+
+    .. code-block:: cpp
+
+        PYBIND11_PLUGIN(example) {
+            pybind11::module_ m("example", "pybind11 example plugin");
+            /// Set up bindings here
+            return m.ptr();
+        }
+\endrst */
+#define PYBIND11_PLUGIN(name)                                                                     \
+    PYBIND11_DEPRECATED("PYBIND11_PLUGIN is deprecated, use PYBIND11_MODULE")                     \
+    static PyObject *pybind11_init();                                                             \
+    PYBIND11_PLUGIN_IMPL(name) {                                                                  \
+        PYBIND11_CHECK_PYTHON_VERSION                                                             \
+        PYBIND11_ENSURE_INTERNALS_READY                                                           \
+        try {                                                                                     \
+            return pybind11_init();                                                               \
+        }                                                                                         \
+        PYBIND11_CATCH_INIT_EXCEPTIONS                                                            \
+    }                                                                                             \
+    PyObject *pybind11_init()
+
+/** \rst
+    This macro creates the entry point that will be invoked when the Python interpreter
+    imports an extension module. The module name is given as the first argument and it
+    should not be in quotes. The second macro argument defines a variable of type
+    `py::module_` which can be used to initialize the module.
+
+    The entry point is marked as "maybe unused" to aid dead-code detection analysis:
+    since the entry point is typically only looked up at runtime and not referenced
+    during translation, it would otherwise appear as unused ("dead") code.
+
+    .. code-block:: cpp
+
+        PYBIND11_MODULE(example, m) {
+            m.doc() = "pybind11 example module";
+
+            // Add bindings here
+            m.def("foo", []() {
+                return "Hello, World!";
+            });
+        }
+
+    The third macro argument is optional (available since 2.13.0), and can be used to
+    mark the extension module as safe to run without the GIL under a free-threaded CPython
+    interpreter. Passing this argument has no effect on other interpreters.
+
+    .. code-block:: cpp
+
+        PYBIND11_MODULE(example, m, py::mod_gil_not_used()) {
+            m.doc() = "pybind11 example module safe to run without the GIL";
+
+            // Add bindings here
+            m.def("foo", []() {
+                return "Hello, Free-threaded World!";
+            });
+        }
+
+\endrst */
+PYBIND11_WARNING_PUSH
+PYBIND11_WARNING_DISABLE_CLANG("-Wgnu-zero-variadic-macro-arguments")
+#define PYBIND11_MODULE(name, variable, ...)                                                      \
+    static ::pybind11::module_::module_def PYBIND11_CONCAT(pybind11_module_def_, name)            \
+        PYBIND11_MAYBE_UNUSED;                                                                    \
+    PYBIND11_MAYBE_UNUSED                                                                         \
+    static void PYBIND11_CONCAT(pybind11_init_, name)(::pybind11::module_ &);                     \
+    PYBIND11_PLUGIN_IMPL(name) {                                                                  \
+        PYBIND11_CHECK_PYTHON_VERSION                                                             \
+        PYBIND11_ENSURE_INTERNALS_READY                                                           \
+        auto m = ::pybind11::module_::create_extension_module(                                    \
+            PYBIND11_TOSTRING(name),                                                              \
+            nullptr,                                                                              \
+            &PYBIND11_CONCAT(pybind11_module_def_, name),                                         \
+            ##__VA_ARGS__);                                                                       \
+        try {                                                                                     \
+            PYBIND11_CONCAT(pybind11_init_, name)(m);                                             \
+            return m.ptr();                                                                       \
+        }                                                                                         \
+        PYBIND11_CATCH_INIT_EXCEPTIONS                                                            \
+    }                                                                                             \
+    void PYBIND11_CONCAT(pybind11_init_, name)(::pybind11::module_ & (variable))
+PYBIND11_WARNING_POP
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+using ssize_t = Py_ssize_t;
+using size_t = std::size_t;
+
+template <typename IntType>
+inline ssize_t ssize_t_cast(const IntType &val) {
+    static_assert(sizeof(IntType) <= sizeof(ssize_t), "Implicit narrowing is not permitted.");
+    return static_cast<ssize_t>(val);
+}
+
+/// Approach used to cast a previously unknown C++ instance into a Python object
+enum class return_value_policy : uint8_t {
+    /** This is the default return value policy, which falls back to the policy
+        return_value_policy::take_ownership when the return value is a pointer.
+        Otherwise, it uses return_value::move or return_value::copy for rvalue
+        and lvalue references, respectively. See below for a description of what
+        all of these different policies do. */
+    automatic = 0,
+
+    /** As above, but use policy return_value_policy::reference when the return
+        value is a pointer. This is the default conversion policy for function
+        arguments when calling Python functions manually from C++ code (i.e. via
+        handle::operator()). You probably won't need to use this. */
+    automatic_reference,
+
+    /** Reference an existing object (i.e. do not create a new copy) and take
+        ownership. Python will call the destructor and delete operator when the
+        object's reference count reaches zero. Undefined behavior ensues when
+        the C++ side does the same.. */
+    take_ownership,
+
+    /** Create a new copy of the returned object, which will be owned by
+        Python. This policy is comparably safe because the lifetimes of the two
+        instances are decoupled. */
+    copy,
+
+    /** Use std::move to move the return value contents into a new instance
+        that will be owned by Python. This policy is comparably safe because the
+        lifetimes of the two instances (move source and destination) are
+        decoupled. */
+    move,
+
+    /** Reference an existing object, but do not take ownership. The C++ side
+        is responsible for managing the object's lifetime and deallocating it
+        when it is no longer used. Warning: undefined behavior will ensue when
+        the C++ side deletes an object that is still referenced and used by
+        Python. */
+    reference,
+
+    /** This policy only applies to methods and properties. It references the
+        object without taking ownership similar to the above
+        return_value_policy::reference policy. In contrast to that policy, the
+        function or property's implicit this argument (called the parent) is
+        considered to be the owner of the return value (the child).
+        pybind11 then couples the lifetime of the parent to the child via a
+        reference relationship that ensures that the parent cannot be garbage
+        collected while Python is still using the child. More advanced
+        variations of this scheme are also possible using combinations of
+        return_value_policy::reference and the keep_alive call policy */
+    reference_internal
+};
+
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+inline static constexpr int log2(size_t n, int k = 0) {
+    return (n <= 1) ? k : log2(n >> 1, k + 1);
+}
+
+// Returns the size as a multiple of sizeof(void *), rounded up.
+inline static constexpr size_t size_in_ptrs(size_t s) {
+    return 1 + ((s - 1) >> log2(sizeof(void *)));
+}
+
+/**
+ * The space to allocate for simple layout instance holders (see below) in multiple of the size of
+ * a pointer (e.g.  2 means 16 bytes on 64-bit architectures).  The default is the minimum required
+ * to holder either a std::unique_ptr or std::shared_ptr (which is almost always
+ * sizeof(std::shared_ptr<T>)).
+ */
+constexpr size_t instance_simple_holder_in_ptrs() {
+    static_assert(sizeof(std::shared_ptr<int>) >= sizeof(std::unique_ptr<int>),
+                  "pybind assumes std::shared_ptrs are at least as big as std::unique_ptrs");
+    return size_in_ptrs(sizeof(std::shared_ptr<int>));
+}
+
+// Forward declarations
+struct type_info;
+struct value_and_holder;
+
+struct nonsimple_values_and_holders {
+    void **values_and_holders;
+    uint8_t *status;
+};
+
+/// The 'instance' type which needs to be standard layout (need to be able to use 'offsetof')
+struct instance {
+    PyObject_HEAD
+    /// Storage for pointers and holder; see simple_layout, below, for a description
+    union {
+        void *simple_value_holder[1 + instance_simple_holder_in_ptrs()];
+        nonsimple_values_and_holders nonsimple;
+    };
+    /// Weak references
+    PyObject *weakrefs;
+    /// If true, the pointer is owned which means we're free to manage it with a holder.
+    bool owned : 1;
+    /**
+     * An instance has two possible value/holder layouts.
+     *
+     * Simple layout (when this flag is true), means the `simple_value_holder` is set with a
+     * pointer and the holder object governing that pointer, i.e. [val1*][holder].  This layout is
+     * applied whenever there is no python-side multiple inheritance of bound C++ types *and* the
+     * type's holder will fit in the default space (which is large enough to hold either a
+     * std::unique_ptr or std::shared_ptr).
+     *
+     * Non-simple layout applies when using custom holders that require more space than
+     * `shared_ptr` (which is typically the size of two pointers), or when multiple inheritance is
+     * used on the python side.  Non-simple layout allocates the required amount of memory to have
+     * multiple bound C++ classes as parents.  Under this layout, `nonsimple.values_and_holders` is
+     * set to a pointer to allocated space of the required space to hold a sequence of value
+     * pointers and holders followed `status`, a set of bit flags (1 byte each), i.e.
+     * [val1*][holder1][val2*][holder2]...[bb...]  where each [block] is rounded up to a multiple
+     * of `sizeof(void *)`.  `nonsimple.status` is, for convenience, a pointer to the beginning of
+     * the [bb...] block (but not independently allocated).
+     *
+     * Status bits indicate whether the associated holder is constructed (&
+     * status_holder_constructed) and whether the value pointer is registered (&
+     * status_instance_registered) in `registered_instances`.
+     */
+    bool simple_layout : 1;
+    /// For simple layout, tracks whether the holder has been constructed
+    bool simple_holder_constructed : 1;
+    /// For simple layout, tracks whether the instance is registered in `registered_instances`
+    bool simple_instance_registered : 1;
+    /// If true, get_internals().patients has an entry for this object
+    bool has_patients : 1;
+
+    /// Initializes all of the above type/values/holders data (but not the instance values
+    /// themselves)
+    void allocate_layout();
+
+    /// Destroys/deallocates all of the above
+    void deallocate_layout();
+
+    /// Returns the value_and_holder wrapper for the given type (or the first, if `find_type`
+    /// omitted).  Returns a default-constructed (with `.inst = nullptr`) object on failure if
+    /// `throw_if_missing` is false.
+    value_and_holder get_value_and_holder(const type_info *find_type = nullptr,
+                                          bool throw_if_missing = true);
+
+    /// Bit values for the non-simple status flags
+    static constexpr uint8_t status_holder_constructed = 1;
+    static constexpr uint8_t status_instance_registered = 2;
+};
+
+static_assert(std::is_standard_layout<instance>::value,
+              "Internal error: `pybind11::detail::instance` is not standard layout!");
+
+/// from __cpp_future__ import (convenient aliases from C++14/17)
+#if defined(PYBIND11_CPP14)
+using std::conditional_t;
+using std::enable_if_t;
+using std::remove_cv_t;
+using std::remove_reference_t;
+#else
+template <bool B, typename T = void>
+using enable_if_t = typename std::enable_if<B, T>::type;
+template <bool B, typename T, typename F>
+using conditional_t = typename std::conditional<B, T, F>::type;
+template <typename T>
+using remove_cv_t = typename std::remove_cv<T>::type;
+template <typename T>
+using remove_reference_t = typename std::remove_reference<T>::type;
+#endif
+
+#if defined(PYBIND11_CPP20)
+using std::remove_cvref;
+using std::remove_cvref_t;
+#else
+template <class T>
+struct remove_cvref {
+    using type = remove_cv_t<remove_reference_t<T>>;
+};
+template <class T>
+using remove_cvref_t = typename remove_cvref<T>::type;
+#endif
+
+/// Example usage: is_same_ignoring_cvref<T, PyObject *>::value
+template <typename T, typename U>
+using is_same_ignoring_cvref = std::is_same<detail::remove_cvref_t<T>, U>;
+
+/// Index sequences
+#if defined(PYBIND11_CPP14)
+using std::index_sequence;
+using std::make_index_sequence;
+#else
+template <size_t...>
+struct index_sequence {};
+template <size_t N, size_t... S>
+struct make_index_sequence_impl : make_index_sequence_impl<N - 1, N - 1, S...> {};
+template <size_t... S>
+struct make_index_sequence_impl<0, S...> {
+    using type = index_sequence<S...>;
+};
+template <size_t N>
+using make_index_sequence = typename make_index_sequence_impl<N>::type;
+#endif
+
+/// Make an index sequence of the indices of true arguments
+template <typename ISeq, size_t, bool...>
+struct select_indices_impl {
+    using type = ISeq;
+};
+template <size_t... IPrev, size_t I, bool B, bool... Bs>
+struct select_indices_impl<index_sequence<IPrev...>, I, B, Bs...>
+    : select_indices_impl<conditional_t<B, index_sequence<IPrev..., I>, index_sequence<IPrev...>>,
+                          I + 1,
+                          Bs...> {};
+template <bool... Bs>
+using select_indices = typename select_indices_impl<index_sequence<>, 0, Bs...>::type;
+
+/// Backports of std::bool_constant and std::negation to accommodate older compilers
+template <bool B>
+using bool_constant = std::integral_constant<bool, B>;
+template <typename T>
+struct negation : bool_constant<!T::value> {};
+
+// PGI/Intel cannot detect operator delete with the "compatible" void_t impl, so
+// using the new one (C++14 defect, so generally works on newer compilers, even
+// if not in C++17 mode)
+#if defined(__PGIC__) || defined(__INTEL_COMPILER)
+template <typename...>
+using void_t = void;
+#else
+template <typename...>
+struct void_t_impl {
+    using type = void;
+};
+template <typename... Ts>
+using void_t = typename void_t_impl<Ts...>::type;
+#endif
+
+/// Compile-time all/any/none of that check the boolean value of all template types
+#if defined(__cpp_fold_expressions) && !(defined(_MSC_VER) && (_MSC_VER < 1916))
+template <class... Ts>
+using all_of = bool_constant<(Ts::value && ...)>;
+template <class... Ts>
+using any_of = bool_constant<(Ts::value || ...)>;
+#elif !defined(_MSC_VER)
+template <bool...>
+struct bools {};
+template <class... Ts>
+using all_of = std::is_same<bools<Ts::value..., true>, bools<true, Ts::value...>>;
+template <class... Ts>
+using any_of = negation<all_of<negation<Ts>...>>;
+#else
+// MSVC has trouble with the above, but supports std::conjunction, which we can use instead (albeit
+// at a slight loss of compilation efficiency).
+template <class... Ts>
+using all_of = std::conjunction<Ts...>;
+template <class... Ts>
+using any_of = std::disjunction<Ts...>;
+#endif
+template <class... Ts>
+using none_of = negation<any_of<Ts...>>;
+
+template <class T, template <class> class... Predicates>
+using satisfies_all_of = all_of<Predicates<T>...>;
+template <class T, template <class> class... Predicates>
+using satisfies_any_of = any_of<Predicates<T>...>;
+template <class T, template <class> class... Predicates>
+using satisfies_none_of = none_of<Predicates<T>...>;
+
+/// Strip the class from a method type
+template <typename T>
+struct remove_class {};
+template <typename C, typename R, typename... A>
+struct remove_class<R (C::*)(A...)> {
+    using type = R(A...);
+};
+template <typename C, typename R, typename... A>
+struct remove_class<R (C::*)(A...) const> {
+    using type = R(A...);
+};
+#ifdef __cpp_noexcept_function_type
+template <typename C, typename R, typename... A>
+struct remove_class<R (C::*)(A...) noexcept> {
+    using type = R(A...);
+};
+template <typename C, typename R, typename... A>
+struct remove_class<R (C::*)(A...) const noexcept> {
+    using type = R(A...);
+};
+#endif
+/// Helper template to strip away type modifiers
+template <typename T>
+struct intrinsic_type {
+    using type = T;
+};
+template <typename T>
+struct intrinsic_type<const T> {
+    using type = typename intrinsic_type<T>::type;
+};
+template <typename T>
+struct intrinsic_type<T *> {
+    using type = typename intrinsic_type<T>::type;
+};
+template <typename T>
+struct intrinsic_type<T &> {
+    using type = typename intrinsic_type<T>::type;
+};
+template <typename T>
+struct intrinsic_type<T &&> {
+    using type = typename intrinsic_type<T>::type;
+};
+template <typename T, size_t N>
+struct intrinsic_type<const T[N]> {
+    using type = typename intrinsic_type<T>::type;
+};
+template <typename T, size_t N>
+struct intrinsic_type<T[N]> {
+    using type = typename intrinsic_type<T>::type;
+};
+template <typename T>
+using intrinsic_t = typename intrinsic_type<T>::type;
+
+/// Helper type to replace 'void' in some expressions
+struct void_type {};
+
+/// Helper template which holds a list of types
+template <typename...>
+struct type_list {};
+
+/// Compile-time integer sum
+#ifdef __cpp_fold_expressions
+template <typename... Ts>
+constexpr size_t constexpr_sum(Ts... ns) {
+    return (0 + ... + size_t{ns});
+}
+#else
+constexpr size_t constexpr_sum() { return 0; }
+template <typename T, typename... Ts>
+constexpr size_t constexpr_sum(T n, Ts... ns) {
+    return size_t{n} + constexpr_sum(ns...);
+}
+#endif
+
+PYBIND11_NAMESPACE_BEGIN(constexpr_impl)
+/// Implementation details for constexpr functions
+constexpr int first(int i) { return i; }
+template <typename T, typename... Ts>
+constexpr int first(int i, T v, Ts... vs) {
+    return v ? i : first(i + 1, vs...);
+}
+
+constexpr int last(int /*i*/, int result) { return result; }
+template <typename T, typename... Ts>
+constexpr int last(int i, int result, T v, Ts... vs) {
+    return last(i + 1, v ? i : result, vs...);
+}
+PYBIND11_NAMESPACE_END(constexpr_impl)
+
+/// Return the index of the first type in Ts which satisfies Predicate<T>.
+/// Returns sizeof...(Ts) if none match.
+template <template <typename> class Predicate, typename... Ts>
+constexpr int constexpr_first() {
+    return constexpr_impl::first(0, Predicate<Ts>::value...);
+}
+
+/// Return the index of the last type in Ts which satisfies Predicate<T>, or -1 if none match.
+template <template <typename> class Predicate, typename... Ts>
+constexpr int constexpr_last() {
+    return constexpr_impl::last(0, -1, Predicate<Ts>::value...);
+}
+
+/// Return the Nth element from the parameter pack
+template <size_t N, typename T, typename... Ts>
+struct pack_element {
+    using type = typename pack_element<N - 1, Ts...>::type;
+};
+template <typename T, typename... Ts>
+struct pack_element<0, T, Ts...> {
+    using type = T;
+};
+
+/// Return the one and only type which matches the predicate, or Default if none match.
+/// If more than one type matches the predicate, fail at compile-time.
+template <template <typename> class Predicate, typename Default, typename... Ts>
+struct exactly_one {
+    static constexpr auto found = constexpr_sum(Predicate<Ts>::value...);
+    static_assert(found <= 1, "Found more than one type matching the predicate");
+
+    static constexpr auto index = found ? constexpr_first<Predicate, Ts...>() : 0;
+    using type = conditional_t<found, typename pack_element<index, Ts...>::type, Default>;
+};
+template <template <typename> class P, typename Default>
+struct exactly_one<P, Default> {
+    using type = Default;
+};
+
+template <template <typename> class Predicate, typename Default, typename... Ts>
+using exactly_one_t = typename exactly_one<Predicate, Default, Ts...>::type;
+
+/// Defer the evaluation of type T until types Us are instantiated
+template <typename T, typename... /*Us*/>
+struct deferred_type {
+    using type = T;
+};
+template <typename T, typename... Us>
+using deferred_t = typename deferred_type<T, Us...>::type;
+
+/// Like is_base_of, but requires a strict base (i.e. `is_strict_base_of<T, T>::value == false`,
+/// unlike `std::is_base_of`)
+template <typename Base, typename Derived>
+using is_strict_base_of
+    = bool_constant<std::is_base_of<Base, Derived>::value && !std::is_same<Base, Derived>::value>;
+
+/// Like is_base_of, but also requires that the base type is accessible (i.e. that a Derived
+/// pointer can be converted to a Base pointer) For unions, `is_base_of<T, T>::value` is False, so
+/// we need to check `is_same` as well.
+template <typename Base, typename Derived>
+using is_accessible_base_of
+    = bool_constant<(std::is_same<Base, Derived>::value || std::is_base_of<Base, Derived>::value)
+                    && std::is_convertible<Derived *, Base *>::value>;
+
+template <template <typename...> class Base>
+struct is_template_base_of_impl {
+    template <typename... Us>
+    static std::true_type check(Base<Us...> *);
+    static std::false_type check(...);
+};
+
+/// Check if a template is the base of a type. For example:
+/// `is_template_base_of<Base, T>` is true if `struct T : Base<U> {}` where U can be anything
+template <template <typename...> class Base, typename T>
+// Sadly, all MSVC versions incl. 2022 need the workaround, even in C++20 mode.
+// See also: https://github.com/pybind/pybind11/pull/3741
+#if !defined(_MSC_VER)
+using is_template_base_of
+    = decltype(is_template_base_of_impl<Base>::check((intrinsic_t<T> *) nullptr));
+#else
+struct is_template_base_of
+    : decltype(is_template_base_of_impl<Base>::check((intrinsic_t<T> *) nullptr)){};
+#endif
+
+/// Check if T is an instantiation of the template `Class`. For example:
+/// `is_instantiation<shared_ptr, T>` is true if `T == shared_ptr<U>` where U can be anything.
+template <template <typename...> class Class, typename T>
+struct is_instantiation : std::false_type {};
+template <template <typename...> class Class, typename... Us>
+struct is_instantiation<Class, Class<Us...>> : std::true_type {};
+
+/// Check if T is std::shared_ptr<U> where U can be anything
+template <typename T>
+using is_shared_ptr = is_instantiation<std::shared_ptr, T>;
+
+/// Check if T looks like an input iterator
+template <typename T, typename = void>
+struct is_input_iterator : std::false_type {};
+template <typename T>
+struct is_input_iterator<T,
+                         void_t<decltype(*std::declval<T &>()), decltype(++std::declval<T &>())>>
+    : std::true_type {};
+
+template <typename T>
+using is_function_pointer
+    = bool_constant<std::is_pointer<T>::value
+                    && std::is_function<typename std::remove_pointer<T>::type>::value>;
+
+template <typename F>
+struct strip_function_object {
+    // If you are encountering an
+    // 'error: name followed by "::" must be a class or namespace name'
+    // with the Intel compiler and a noexcept function here,
+    // try to use noexcept(true) instead of plain noexcept.
+    using type = typename remove_class<decltype(&F::operator())>::type;
+};
+
+// Extracts the function signature from a function, function pointer or lambda.
+template <typename Function, typename F = remove_reference_t<Function>>
+using function_signature_t = conditional_t<
+    std::is_function<F>::value,
+    F,
+    typename conditional_t<std::is_pointer<F>::value || std::is_member_pointer<F>::value,
+                           std::remove_pointer<F>,
+                           strip_function_object<F>>::type>;
+
+/// Returns true if the type looks like a lambda: that is, isn't a function, pointer or member
+/// pointer.  Note that this can catch all sorts of other things, too; this is intended to be used
+/// in a place where passing a lambda makes sense.
+template <typename T>
+using is_lambda = satisfies_none_of<remove_reference_t<T>,
+                                    std::is_function,
+                                    std::is_pointer,
+                                    std::is_member_pointer>;
+
+// [workaround(intel)] Internal error on fold expression
+/// Apply a function over each element of a parameter pack
+#if defined(__cpp_fold_expressions) && !defined(__INTEL_COMPILER)
+// Intel compiler produces an internal error on this fold expression (tested with ICC 19.0.2)
+#    define PYBIND11_EXPAND_SIDE_EFFECTS(PATTERN) (((PATTERN), void()), ...)
+#else
+using expand_side_effects = bool[];
+#    define PYBIND11_EXPAND_SIDE_EFFECTS(PATTERN)                                                 \
+        (void) pybind11::detail::expand_side_effects { ((PATTERN), void(), false)..., false }
+#endif
+
+PYBIND11_NAMESPACE_END(detail)
+
+/// C++ bindings of builtin Python exceptions
+class PYBIND11_EXPORT_EXCEPTION builtin_exception : public std::runtime_error {
+public:
+    using std::runtime_error::runtime_error;
+    /// Set the error using the Python C API
+    virtual void set_error() const = 0;
+};
+
+#define PYBIND11_RUNTIME_EXCEPTION(name, type)                                                    \
+    class PYBIND11_EXPORT_EXCEPTION name : public builtin_exception {                             \
+    public:                                                                                       \
+        using builtin_exception::builtin_exception;                                               \
+        name() : name("") {}                                                                      \
+        void set_error() const override { PyErr_SetString(type, what()); }                        \
+    };
+
+PYBIND11_RUNTIME_EXCEPTION(stop_iteration, PyExc_StopIteration)
+PYBIND11_RUNTIME_EXCEPTION(index_error, PyExc_IndexError)
+PYBIND11_RUNTIME_EXCEPTION(key_error, PyExc_KeyError)
+PYBIND11_RUNTIME_EXCEPTION(value_error, PyExc_ValueError)
+PYBIND11_RUNTIME_EXCEPTION(type_error, PyExc_TypeError)
+PYBIND11_RUNTIME_EXCEPTION(buffer_error, PyExc_BufferError)
+PYBIND11_RUNTIME_EXCEPTION(import_error, PyExc_ImportError)
+PYBIND11_RUNTIME_EXCEPTION(attribute_error, PyExc_AttributeError)
+PYBIND11_RUNTIME_EXCEPTION(cast_error, PyExc_RuntimeError) /// Thrown when pybind11::cast or
+                                                           /// handle::call fail due to a type
+                                                           /// casting error
+PYBIND11_RUNTIME_EXCEPTION(reference_cast_error, PyExc_RuntimeError) /// Used internally
+
+[[noreturn]] PYBIND11_NOINLINE void pybind11_fail(const char *reason) {
+    assert(!PyErr_Occurred());
+    throw std::runtime_error(reason);
+}
+[[noreturn]] PYBIND11_NOINLINE void pybind11_fail(const std::string &reason) {
+    assert(!PyErr_Occurred());
+    throw std::runtime_error(reason);
+}
+
+template <typename T, typename SFINAE = void>
+struct format_descriptor {};
+
+template <typename T>
+struct format_descriptor<
+    T,
+    detail::enable_if_t<detail::is_same_ignoring_cvref<T, PyObject *>::value>> {
+    static constexpr const char c = 'O';
+    static constexpr const char value[2] = {c, '\0'};
+    static std::string format() { return std::string(1, c); }
+};
+
+PYBIND11_NAMESPACE_BEGIN(detail)
+// Returns the index of the given type in the type char array below, and in the list in numpy.h
+// The order here is: bool; 8 ints ((signed,unsigned)x(8,16,32,64)bits); float,double,long double;
+// complex float,double,long double.  Note that the long double types only participate when long
+// double is actually longer than double (it isn't under MSVC).
+// NB: not only the string below but also complex.h and numpy.h rely on this order.
+template <typename T, typename SFINAE = void>
+struct is_fmt_numeric {
+    static constexpr bool value = false;
+};
+template <typename T>
+struct is_fmt_numeric<T, enable_if_t<std::is_arithmetic<T>::value>> {
+    static constexpr bool value = true;
+    static constexpr int index
+        = std::is_same<T, bool>::value
+              ? 0
+              : 1
+                    + (std::is_integral<T>::value
+                           ? detail::log2(sizeof(T)) * 2 + std::is_unsigned<T>::value
+                           : 8
+                                 + (std::is_same<T, double>::value        ? 1
+                                    : std::is_same<T, long double>::value ? 2
+                                                                          : 0));
+};
+PYBIND11_NAMESPACE_END(detail)
+
+template <typename T>
+struct format_descriptor<T, detail::enable_if_t<std::is_arithmetic<T>::value>> {
+    static constexpr const char c = "?bBhHiIqQfdg"[detail::is_fmt_numeric<T>::index];
+    static constexpr const char value[2] = {c, '\0'};
+    static std::string format() { return std::string(1, c); }
+};
+
+#if !defined(PYBIND11_CPP17)
+
+template <typename T>
+constexpr const char
+    format_descriptor<T, detail::enable_if_t<std::is_arithmetic<T>::value>>::value[2];
+
+#endif
+
+/// RAII wrapper that temporarily clears any Python error state
+struct error_scope {
+    PyObject *type, *value, *trace;
+    error_scope() { PyErr_Fetch(&type, &value, &trace); }
+    error_scope(const error_scope &) = delete;
+    error_scope &operator=(const error_scope &) = delete;
+    ~error_scope() { PyErr_Restore(type, value, trace); }
+};
+
+/// Dummy destructor wrapper that can be used to expose classes with a private destructor
+struct nodelete {
+    template <typename T>
+    void operator()(T *) {}
+};
+
+PYBIND11_NAMESPACE_BEGIN(detail)
+template <typename... Args>
+struct overload_cast_impl {
+    template <typename Return>
+    constexpr auto operator()(Return (*pf)(Args...)) const noexcept -> decltype(pf) {
+        return pf;
+    }
+
+    template <typename Return, typename Class>
+    constexpr auto operator()(Return (Class::*pmf)(Args...),
+                              std::false_type = {}) const noexcept -> decltype(pmf) {
+        return pmf;
+    }
+
+    template <typename Return, typename Class>
+    constexpr auto operator()(Return (Class::*pmf)(Args...) const,
+                              std::true_type) const noexcept -> decltype(pmf) {
+        return pmf;
+    }
+};
+PYBIND11_NAMESPACE_END(detail)
+
+// overload_cast requires variable templates: C++14
+#if defined(PYBIND11_CPP14)
+#    define PYBIND11_OVERLOAD_CAST 1
+/// Syntax sugar for resolving overloaded function pointers:
+///  - regular: static_cast<Return (Class::*)(Arg0, Arg1, Arg2)>(&Class::func)
+///  - sweet:   overload_cast<Arg0, Arg1, Arg2>(&Class::func)
+template <typename... Args>
+static constexpr detail::overload_cast_impl<Args...> overload_cast{};
+#endif
+
+/// Const member function selector for overload_cast
+///  - regular: static_cast<Return (Class::*)(Arg) const>(&Class::func)
+///  - sweet:   overload_cast<Arg>(&Class::func, const_)
+static constexpr auto const_ = std::true_type{};
+
+#if !defined(PYBIND11_CPP14) // no overload_cast: providing something that static_assert-fails:
+template <typename... Args>
+struct overload_cast {
+    static_assert(detail::deferred_t<std::false_type, Args...>::value,
+                  "pybind11::overload_cast<...> requires compiling in C++14 mode");
+};
+#endif // overload_cast
+
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+// Adaptor for converting arbitrary container arguments into a vector; implicitly convertible from
+// any standard container (or C-style array) supporting std::begin/std::end, any singleton
+// arithmetic type (if T is arithmetic), or explicitly constructible from an iterator pair.
+template <typename T>
+class any_container {
+    std::vector<T> v;
+
+public:
+    any_container() = default;
+
+    // Can construct from a pair of iterators
+    template <typename It, typename = enable_if_t<is_input_iterator<It>::value>>
+    any_container(It first, It last) : v(first, last) {}
+
+    // Implicit conversion constructor from any arbitrary container type
+    // with values convertible to T
+    template <typename Container,
+              typename = enable_if_t<
+                  std::is_convertible<decltype(*std::begin(std::declval<const Container &>())),
+                                      T>::value>>
+    // NOLINTNEXTLINE(google-explicit-constructor)
+    any_container(const Container &c) : any_container(std::begin(c), std::end(c)) {}
+
+    // initializer_list's aren't deducible, so don't get matched by the above template;
+    // we need this to explicitly allow implicit conversion from one:
+    template <typename TIn, typename = enable_if_t<std::is_convertible<TIn, T>::value>>
+    any_container(const std::initializer_list<TIn> &c) : any_container(c.begin(), c.end()) {}
+
+    // Avoid copying if given an rvalue vector of the correct type.
+    // NOLINTNEXTLINE(google-explicit-constructor)
+    any_container(std::vector<T> &&v) : v(std::move(v)) {}
+
+    // Moves the vector out of an rvalue any_container
+    // NOLINTNEXTLINE(google-explicit-constructor)
+    operator std::vector<T> &&() && { return std::move(v); }
+
+    // Dereferencing obtains a reference to the underlying vector
+    std::vector<T> &operator*() { return v; }
+    const std::vector<T> &operator*() const { return v; }
+
+    // -> lets you call methods on the underlying vector
+    std::vector<T> *operator->() { return &v; }
+    const std::vector<T> *operator->() const { return &v; }
+};
+
+// Forward-declaration; see detail/class.h
+std::string get_fully_qualified_tp_name(PyTypeObject *);
+
+template <typename T>
+inline static std::shared_ptr<T>
+try_get_shared_from_this(std::enable_shared_from_this<T> *holder_value_ptr) {
+// Pre C++17, this code path exploits undefined behavior, but is known to work on many platforms.
+// Use at your own risk!
+// See also https://en.cppreference.com/w/cpp/memory/enable_shared_from_this, and in particular
+// the `std::shared_ptr<Good> gp1 = not_so_good.getptr();` and `try`-`catch` parts of the example.
+#if defined(__cpp_lib_enable_shared_from_this) && (!defined(_MSC_VER) || _MSC_VER >= 1912)
+    return holder_value_ptr->weak_from_this().lock();
+#else
+    try {
+        return holder_value_ptr->shared_from_this();
+    } catch (const std::bad_weak_ptr &) {
+        return nullptr;
+    }
+#endif
+}
+
+// For silencing "unused" compiler warnings in special situations.
+template <typename... Args>
+#if defined(_MSC_VER) && _MSC_VER < 1920 // MSVC 2017
+constexpr
+#endif
+    inline void
+    silence_unused_warnings(Args &&...) {
+}
+
+// MSVC warning C4100: Unreferenced formal parameter
+#if defined(_MSC_VER) && _MSC_VER <= 1916
+#    define PYBIND11_WORKAROUND_INCORRECT_MSVC_C4100(...)                                         \
+        detail::silence_unused_warnings(__VA_ARGS__)
+#else
+#    define PYBIND11_WORKAROUND_INCORRECT_MSVC_C4100(...)
+#endif
+
+// GCC -Wunused-but-set-parameter  All GCC versions (as of July 2021).
+#if defined(__GNUG__) && !defined(__clang__) && !defined(__INTEL_COMPILER)
+#    define PYBIND11_WORKAROUND_INCORRECT_GCC_UNUSED_BUT_SET_PARAMETER(...)                       \
+        detail::silence_unused_warnings(__VA_ARGS__)
+#else
+#    define PYBIND11_WORKAROUND_INCORRECT_GCC_UNUSED_BUT_SET_PARAMETER(...)
+#endif
+
+#if defined(__clang__)                                                                            \
+    && (defined(__apple_build_version__) /* AppleClang 13.0.0.13000029 was the only data point    \
+                                            available. */                                         \
+        || (__clang_major__ >= 7                                                                  \
+            && __clang_major__ <= 12) /* Clang 3, 5, 13, 14, 15 do not generate the warning. */   \
+    )
+#    define PYBIND11_DETECTED_CLANG_WITH_MISLEADING_CALL_STD_MOVE_EXPLICITLY_WARNING
+// Example:
+// tests/test_kwargs_and_defaults.cpp:46:68: error: local variable 'args' will be copied despite
+// being returned by name [-Werror,-Wreturn-std-move]
+//     m.def("args_function", [](py::args args) -> py::tuple { return args; });
+//                                                                    ^~~~
+// test_kwargs_and_defaults.cpp:46:68: note: call 'std::move' explicitly to avoid copying
+//     m.def("args_function", [](py::args args) -> py::tuple { return args; });
+//                                                                    ^~~~
+//                                                                    std::move(args)
+#endif
+
+// Pybind offers detailed error messages by default for all builts that are debug (through the
+// negation of NDEBUG). This can also be manually enabled by users, for any builds, through
+// defining PYBIND11_DETAILED_ERROR_MESSAGES. This information is primarily useful for those
+// who are writing (as opposed to merely using) libraries that use pybind11.
+#if !defined(PYBIND11_DETAILED_ERROR_MESSAGES) && !defined(NDEBUG)
+#    define PYBIND11_DETAILED_ERROR_MESSAGES
+#endif
+
+PYBIND11_NAMESPACE_END(detail)
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/detail/cpp_conduit.h

@@ -0,0 +1,77 @@
+// Copyright (c) 2024 The pybind Community.
+
+#pragma once
+
+#include <pybind11/pytypes.h>
+
+#include "common.h"
+#include "internals.h"
+
+#include <typeinfo>
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+// Forward declaration needed here: Refactoring opportunity.
+extern "C" inline PyObject *pybind11_object_new(PyTypeObject *type, PyObject *, PyObject *);
+
+inline bool type_is_managed_by_our_internals(PyTypeObject *type_obj) {
+#if defined(PYPY_VERSION)
+    auto &internals = get_internals();
+    return bool(internals.registered_types_py.find(type_obj)
+                != internals.registered_types_py.end());
+#else
+    return bool(type_obj->tp_new == pybind11_object_new);
+#endif
+}
+
+inline bool is_instance_method_of_type(PyTypeObject *type_obj, PyObject *attr_name) {
+    PyObject *descr = _PyType_Lookup(type_obj, attr_name);
+    return bool((descr != nullptr) && PyInstanceMethod_Check(descr));
+}
+
+inline object try_get_cpp_conduit_method(PyObject *obj) {
+    if (PyType_Check(obj)) {
+        return object();
+    }
+    PyTypeObject *type_obj = Py_TYPE(obj);
+    str attr_name("_pybind11_conduit_v1_");
+    bool assumed_to_be_callable = false;
+    if (type_is_managed_by_our_internals(type_obj)) {
+        if (!is_instance_method_of_type(type_obj, attr_name.ptr())) {
+            return object();
+        }
+        assumed_to_be_callable = true;
+    }
+    PyObject *method = PyObject_GetAttr(obj, attr_name.ptr());
+    if (method == nullptr) {
+        PyErr_Clear();
+        return object();
+    }
+    if (!assumed_to_be_callable && PyCallable_Check(method) == 0) {
+        Py_DECREF(method);
+        return object();
+    }
+    return reinterpret_steal<object>(method);
+}
+
+inline void *try_raw_pointer_ephemeral_from_cpp_conduit(handle src,
+                                                        const std::type_info *cpp_type_info) {
+    object method = try_get_cpp_conduit_method(src.ptr());
+    if (method) {
+        capsule cpp_type_info_capsule(const_cast<void *>(static_cast<const void *>(cpp_type_info)),
+                                      typeid(std::type_info).name());
+        object cpp_conduit = method(bytes(PYBIND11_PLATFORM_ABI_ID),
+                                    cpp_type_info_capsule,
+                                    bytes("raw_pointer_ephemeral"));
+        if (isinstance<capsule>(cpp_conduit)) {
+            return reinterpret_borrow<capsule>(cpp_conduit).get_pointer();
+        }
+    }
+    return nullptr;
+}
+
+#define PYBIND11_HAS_CPP_CONDUIT 1
+
+PYBIND11_NAMESPACE_END(detail)
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/detail/descr.h

@@ -0,0 +1,172 @@
+/*
+    pybind11/detail/descr.h: Helper type for concatenating type signatures at compile time
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "common.h"
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+#if !defined(_MSC_VER)
+#    define PYBIND11_DESCR_CONSTEXPR static constexpr
+#else
+#    define PYBIND11_DESCR_CONSTEXPR const
+#endif
+
+/* Concatenate type signatures at compile time */
+template <size_t N, typename... Ts>
+struct descr {
+    char text[N + 1]{'\0'};
+
+    constexpr descr() = default;
+    // NOLINTNEXTLINE(google-explicit-constructor)
+    constexpr descr(char const (&s)[N + 1]) : descr(s, make_index_sequence<N>()) {}
+
+    template <size_t... Is>
+    constexpr descr(char const (&s)[N + 1], index_sequence<Is...>) : text{s[Is]..., '\0'} {}
+
+    template <typename... Chars>
+    // NOLINTNEXTLINE(google-explicit-constructor)
+    constexpr descr(char c, Chars... cs) : text{c, static_cast<char>(cs)..., '\0'} {}
+
+    static constexpr std::array<const std::type_info *, sizeof...(Ts) + 1> types() {
+        return {{&typeid(Ts)..., nullptr}};
+    }
+};
+
+template <size_t N1, size_t N2, typename... Ts1, typename... Ts2, size_t... Is1, size_t... Is2>
+constexpr descr<N1 + N2, Ts1..., Ts2...> plus_impl(const descr<N1, Ts1...> &a,
+                                                   const descr<N2, Ts2...> &b,
+                                                   index_sequence<Is1...>,
+                                                   index_sequence<Is2...>) {
+    PYBIND11_WORKAROUND_INCORRECT_MSVC_C4100(b);
+    return {a.text[Is1]..., b.text[Is2]...};
+}
+
+template <size_t N1, size_t N2, typename... Ts1, typename... Ts2>
+constexpr descr<N1 + N2, Ts1..., Ts2...> operator+(const descr<N1, Ts1...> &a,
+                                                   const descr<N2, Ts2...> &b) {
+    return plus_impl(a, b, make_index_sequence<N1>(), make_index_sequence<N2>());
+}
+
+template <size_t N>
+constexpr descr<N - 1> const_name(char const (&text)[N]) {
+    return descr<N - 1>(text);
+}
+constexpr descr<0> const_name(char const (&)[1]) { return {}; }
+
+template <size_t Rem, size_t... Digits>
+struct int_to_str : int_to_str<Rem / 10, Rem % 10, Digits...> {};
+template <size_t... Digits>
+struct int_to_str<0, Digits...> {
+    // WARNING: This only works with C++17 or higher.
+    static constexpr auto digits = descr<sizeof...(Digits)>(('0' + Digits)...);
+};
+
+// Ternary description (like std::conditional)
+template <bool B, size_t N1, size_t N2>
+constexpr enable_if_t<B, descr<N1 - 1>> const_name(char const (&text1)[N1], char const (&)[N2]) {
+    return const_name(text1);
+}
+template <bool B, size_t N1, size_t N2>
+constexpr enable_if_t<!B, descr<N2 - 1>> const_name(char const (&)[N1], char const (&text2)[N2]) {
+    return const_name(text2);
+}
+
+template <bool B, typename T1, typename T2>
+constexpr enable_if_t<B, T1> const_name(const T1 &d, const T2 &) {
+    return d;
+}
+template <bool B, typename T1, typename T2>
+constexpr enable_if_t<!B, T2> const_name(const T1 &, const T2 &d) {
+    return d;
+}
+
+template <size_t Size>
+auto constexpr const_name() -> remove_cv_t<decltype(int_to_str<Size / 10, Size % 10>::digits)> {
+    return int_to_str<Size / 10, Size % 10>::digits;
+}
+
+template <typename Type>
+constexpr descr<1, Type> const_name() {
+    return {'%'};
+}
+
+// If "_" is defined as a macro, py::detail::_ cannot be provided.
+// It is therefore best to use py::detail::const_name universally.
+// This block is for backward compatibility only.
+// (The const_name code is repeated to avoid introducing a "_" #define ourselves.)
+#ifndef _
+#    define PYBIND11_DETAIL_UNDERSCORE_BACKWARD_COMPATIBILITY
+template <size_t N>
+constexpr descr<N - 1> _(char const (&text)[N]) {
+    return const_name<N>(text);
+}
+template <bool B, size_t N1, size_t N2>
+constexpr enable_if_t<B, descr<N1 - 1>> _(char const (&text1)[N1], char const (&text2)[N2]) {
+    return const_name<B, N1, N2>(text1, text2);
+}
+template <bool B, size_t N1, size_t N2>
+constexpr enable_if_t<!B, descr<N2 - 1>> _(char const (&text1)[N1], char const (&text2)[N2]) {
+    return const_name<B, N1, N2>(text1, text2);
+}
+template <bool B, typename T1, typename T2>
+constexpr enable_if_t<B, T1> _(const T1 &d1, const T2 &d2) {
+    return const_name<B, T1, T2>(d1, d2);
+}
+template <bool B, typename T1, typename T2>
+constexpr enable_if_t<!B, T2> _(const T1 &d1, const T2 &d2) {
+    return const_name<B, T1, T2>(d1, d2);
+}
+
+template <size_t Size>
+auto constexpr _() -> remove_cv_t<decltype(int_to_str<Size / 10, Size % 10>::digits)> {
+    return const_name<Size>();
+}
+template <typename Type>
+constexpr descr<1, Type> _() {
+    return const_name<Type>();
+}
+#endif // #ifndef _
+
+constexpr descr<0> concat() { return {}; }
+
+template <size_t N, typename... Ts>
+constexpr descr<N, Ts...> concat(const descr<N, Ts...> &descr) {
+    return descr;
+}
+
+#ifdef __cpp_fold_expressions
+template <size_t N1, size_t N2, typename... Ts1, typename... Ts2>
+constexpr descr<N1 + N2 + 2, Ts1..., Ts2...> operator,(const descr<N1, Ts1...> &a,
+                                                       const descr<N2, Ts2...> &b) {
+    return a + const_name(", ") + b;
+}
+
+template <size_t N, typename... Ts, typename... Args>
+constexpr auto concat(const descr<N, Ts...> &d, const Args &...args) {
+    return (d, ..., args);
+}
+#else
+template <size_t N, typename... Ts, typename... Args>
+constexpr auto concat(const descr<N, Ts...> &d,
+                      const Args &...args) -> decltype(std::declval<descr<N + 2, Ts...>>()
+                                                       + concat(args...)) {
+    return d + const_name(", ") + concat(args...);
+}
+#endif
+
+template <size_t N, typename... Ts>
+constexpr descr<N + 2, Ts...> type_descr(const descr<N, Ts...> &descr) {
+    return const_name("{") + descr + const_name("}");
+}
+
+PYBIND11_NAMESPACE_END(detail)
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/detail/exception_translation.h

@@ -0,0 +1,71 @@
+/*
+    pybind11/detail/exception_translation.h: means to translate C++ exceptions to Python exceptions
+
+    Copyright (c) 2024 The Pybind Development Team.
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "common.h"
+#include "internals.h"
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+// Apply all the extensions translators from a list
+// Return true if one of the translators completed without raising an exception
+// itself. Return of false indicates that if there are other translators
+// available, they should be tried.
+inline bool apply_exception_translators(std::forward_list<ExceptionTranslator> &translators) {
+    auto last_exception = std::current_exception();
+
+    for (auto &translator : translators) {
+        try {
+            translator(last_exception);
+            return true;
+        } catch (...) {
+            last_exception = std::current_exception();
+        }
+    }
+    return false;
+}
+
+inline void try_translate_exceptions() {
+    /* When an exception is caught, give each registered exception
+        translator a chance to translate it to a Python exception. First
+        all module-local translators will be tried in reverse order of
+        registration. If none of the module-locale translators handle
+        the exception (or there are no module-locale translators) then
+        the global translators will be tried, also in reverse order of
+        registration.
+
+        A translator may choose to do one of the following:
+
+        - catch the exception and call py::set_error()
+            to set a standard (or custom) Python exception, or
+        - do nothing and let the exception fall through to the next translator, or
+        - delegate translation to the next translator by throwing a new type of exception.
+        */
+
+    bool handled = with_internals([&](internals &internals) {
+        auto &local_exception_translators = get_local_internals().registered_exception_translators;
+        if (detail::apply_exception_translators(local_exception_translators)) {
+            return true;
+        }
+        auto &exception_translators = internals.registered_exception_translators;
+        if (detail::apply_exception_translators(exception_translators)) {
+            return true;
+        }
+        return false;
+    });
+
+    if (!handled) {
+        set_error(PyExc_SystemError, "Exception escaped from default exception translator!");
+    }
+}
+
+PYBIND11_NAMESPACE_END(detail)
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/detail/init.h

@@ -0,0 +1,436 @@
+/*
+    pybind11/detail/init.h: init factory function implementation and support code.
+
+    Copyright (c) 2017 Jason Rhinelander <jason@imaginary.ca>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "class.h"
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+PYBIND11_WARNING_DISABLE_MSVC(4127)
+
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+template <>
+class type_caster<value_and_holder> {
+public:
+    bool load(handle h, bool) {
+        value = reinterpret_cast<value_and_holder *>(h.ptr());
+        return true;
+    }
+
+    template <typename>
+    using cast_op_type = value_and_holder &;
+    explicit operator value_and_holder &() { return *value; }
+    static constexpr auto name = const_name<value_and_holder>();
+
+private:
+    value_and_holder *value = nullptr;
+};
+
+PYBIND11_NAMESPACE_BEGIN(initimpl)
+
+inline void no_nullptr(void *ptr) {
+    if (!ptr) {
+        throw type_error("pybind11::init(): factory function returned nullptr");
+    }
+}
+
+// Implementing functions for all forms of py::init<...> and py::init(...)
+template <typename Class>
+using Cpp = typename Class::type;
+template <typename Class>
+using Alias = typename Class::type_alias;
+template <typename Class>
+using Holder = typename Class::holder_type;
+
+template <typename Class>
+using is_alias_constructible = std::is_constructible<Alias<Class>, Cpp<Class> &&>;
+
+// Takes a Cpp pointer and returns true if it actually is a polymorphic Alias instance.
+template <typename Class, enable_if_t<Class::has_alias, int> = 0>
+bool is_alias(Cpp<Class> *ptr) {
+    return dynamic_cast<Alias<Class> *>(ptr) != nullptr;
+}
+// Failing fallback version of the above for a no-alias class (always returns false)
+template <typename /*Class*/>
+constexpr bool is_alias(void *) {
+    return false;
+}
+
+// Constructs and returns a new object; if the given arguments don't map to a constructor, we fall
+// back to brace aggregate initialization so that for aggregate initialization can be used with
+// py::init, e.g.  `py::init<int, int>` to initialize a `struct T { int a; int b; }`.  For
+// non-aggregate types, we need to use an ordinary T(...) constructor (invoking as `T{...}` usually
+// works, but will not do the expected thing when `T` has an `initializer_list<T>` constructor).
+template <typename Class,
+          typename... Args,
+          detail::enable_if_t<std::is_constructible<Class, Args...>::value, int> = 0>
+inline Class *construct_or_initialize(Args &&...args) {
+    return new Class(std::forward<Args>(args)...);
+}
+template <typename Class,
+          typename... Args,
+          detail::enable_if_t<!std::is_constructible<Class, Args...>::value, int> = 0>
+inline Class *construct_or_initialize(Args &&...args) {
+    return new Class{std::forward<Args>(args)...};
+}
+
+// Attempts to constructs an alias using a `Alias(Cpp &&)` constructor.  This allows types with
+// an alias to provide only a single Cpp factory function as long as the Alias can be
+// constructed from an rvalue reference of the base Cpp type.  This means that Alias classes
+// can, when appropriate, simply define a `Alias(Cpp &&)` constructor rather than needing to
+// inherit all the base class constructors.
+template <typename Class>
+void construct_alias_from_cpp(std::true_type /*is_alias_constructible*/,
+                              value_and_holder &v_h,
+                              Cpp<Class> &&base) {
+    v_h.value_ptr() = new Alias<Class>(std::move(base));
+}
+template <typename Class>
+[[noreturn]] void construct_alias_from_cpp(std::false_type /*!is_alias_constructible*/,
+                                           value_and_holder &,
+                                           Cpp<Class> &&) {
+    throw type_error("pybind11::init(): unable to convert returned instance to required "
+                     "alias class: no `Alias<Class>(Class &&)` constructor available");
+}
+
+// Error-generating fallback for factories that don't match one of the below construction
+// mechanisms.
+template <typename Class>
+void construct(...) {
+    static_assert(!std::is_same<Class, Class>::value /* always false */,
+                  "pybind11::init(): init function must return a compatible pointer, "
+                  "holder, or value");
+}
+
+// Pointer return v1: the factory function returns a class pointer for a registered class.
+// If we don't need an alias (because this class doesn't have one, or because the final type is
+// inherited on the Python side) we can simply take over ownership.  Otherwise we need to try to
+// construct an Alias from the returned base instance.
+template <typename Class>
+void construct(value_and_holder &v_h, Cpp<Class> *ptr, bool need_alias) {
+    PYBIND11_WORKAROUND_INCORRECT_MSVC_C4100(need_alias);
+    no_nullptr(ptr);
+    if (Class::has_alias && need_alias && !is_alias<Class>(ptr)) {
+        // We're going to try to construct an alias by moving the cpp type.  Whether or not
+        // that succeeds, we still need to destroy the original cpp pointer (either the
+        // moved away leftover, if the alias construction works, or the value itself if we
+        // throw an error), but we can't just call `delete ptr`: it might have a special
+        // deleter, or might be shared_from_this.  So we construct a holder around it as if
+        // it was a normal instance, then steal the holder away into a local variable; thus
+        // the holder and destruction happens when we leave the C++ scope, and the holder
+        // class gets to handle the destruction however it likes.
+        v_h.value_ptr() = ptr;
+        v_h.set_instance_registered(true); // Trick to prevent init_instance from registering it
+        // DANGER ZONE BEGIN: exceptions will leave v_h in an invalid state.
+        v_h.type->init_instance(v_h.inst, nullptr);                        // Set up the holder
+        Holder<Class> temp_holder(std::move(v_h.holder<Holder<Class>>())); // Steal the holder
+        v_h.type->dealloc(v_h); // Destroys the moved-out holder remains, resets value ptr to null
+        v_h.set_instance_registered(false);
+        // DANGER ZONE END.
+
+        construct_alias_from_cpp<Class>(is_alias_constructible<Class>{}, v_h, std::move(*ptr));
+    } else {
+        // Otherwise the type isn't inherited, so we don't need an Alias
+        v_h.value_ptr() = ptr;
+    }
+}
+
+// Pointer return v2: a factory that always returns an alias instance ptr.  We simply take over
+// ownership of the pointer.
+template <typename Class, enable_if_t<Class::has_alias, int> = 0>
+void construct(value_and_holder &v_h, Alias<Class> *alias_ptr, bool) {
+    no_nullptr(alias_ptr);
+    v_h.value_ptr() = static_cast<Cpp<Class> *>(alias_ptr);
+}
+
+// Holder return: copy its pointer, and move or copy the returned holder into the new instance's
+// holder.  This also handles types like std::shared_ptr<T> and std::unique_ptr<T> where T is a
+// derived type (through those holder's implicit conversion from derived class holder
+// constructors).
+template <typename Class>
+void construct(value_and_holder &v_h, Holder<Class> holder, bool need_alias) {
+    PYBIND11_WORKAROUND_INCORRECT_MSVC_C4100(need_alias);
+    auto *ptr = holder_helper<Holder<Class>>::get(holder);
+    no_nullptr(ptr);
+    // If we need an alias, check that the held pointer is actually an alias instance
+    if (Class::has_alias && need_alias && !is_alias<Class>(ptr)) {
+        throw type_error("pybind11::init(): construction failed: returned holder-wrapped instance "
+                         "is not an alias instance");
+    }
+
+    v_h.value_ptr() = ptr;
+    v_h.type->init_instance(v_h.inst, &holder);
+}
+
+// return-by-value version 1: returning a cpp class by value.  If the class has an alias and an
+// alias is required the alias must have an `Alias(Cpp &&)` constructor so that we can construct
+// the alias from the base when needed (i.e. because of Python-side inheritance).  When we don't
+// need it, we simply move-construct the cpp value into a new instance.
+template <typename Class>
+void construct(value_and_holder &v_h, Cpp<Class> &&result, bool need_alias) {
+    PYBIND11_WORKAROUND_INCORRECT_MSVC_C4100(need_alias);
+    static_assert(is_move_constructible<Cpp<Class>>::value,
+                  "pybind11::init() return-by-value factory function requires a movable class");
+    if (Class::has_alias && need_alias) {
+        construct_alias_from_cpp<Class>(is_alias_constructible<Class>{}, v_h, std::move(result));
+    } else {
+        v_h.value_ptr() = new Cpp<Class>(std::move(result));
+    }
+}
+
+// return-by-value version 2: returning a value of the alias type itself.  We move-construct an
+// Alias instance (even if no the python-side inheritance is involved).  The is intended for
+// cases where Alias initialization is always desired.
+template <typename Class>
+void construct(value_and_holder &v_h, Alias<Class> &&result, bool) {
+    static_assert(
+        is_move_constructible<Alias<Class>>::value,
+        "pybind11::init() return-by-alias-value factory function requires a movable alias class");
+    v_h.value_ptr() = new Alias<Class>(std::move(result));
+}
+
+// Implementing class for py::init<...>()
+template <typename... Args>
+struct constructor {
+    template <typename Class, typename... Extra, enable_if_t<!Class::has_alias, int> = 0>
+    static void execute(Class &cl, const Extra &...extra) {
+        cl.def(
+            "__init__",
+            [](value_and_holder &v_h, Args... args) {
+                v_h.value_ptr() = construct_or_initialize<Cpp<Class>>(std::forward<Args>(args)...);
+            },
+            is_new_style_constructor(),
+            extra...);
+    }
+
+    template <
+        typename Class,
+        typename... Extra,
+        enable_if_t<Class::has_alias && std::is_constructible<Cpp<Class>, Args...>::value, int>
+        = 0>
+    static void execute(Class &cl, const Extra &...extra) {
+        cl.def(
+            "__init__",
+            [](value_and_holder &v_h, Args... args) {
+                if (Py_TYPE(v_h.inst) == v_h.type->type) {
+                    v_h.value_ptr()
+                        = construct_or_initialize<Cpp<Class>>(std::forward<Args>(args)...);
+                } else {
+                    v_h.value_ptr()
+                        = construct_or_initialize<Alias<Class>>(std::forward<Args>(args)...);
+                }
+            },
+            is_new_style_constructor(),
+            extra...);
+    }
+
+    template <
+        typename Class,
+        typename... Extra,
+        enable_if_t<Class::has_alias && !std::is_constructible<Cpp<Class>, Args...>::value, int>
+        = 0>
+    static void execute(Class &cl, const Extra &...extra) {
+        cl.def(
+            "__init__",
+            [](value_and_holder &v_h, Args... args) {
+                v_h.value_ptr()
+                    = construct_or_initialize<Alias<Class>>(std::forward<Args>(args)...);
+            },
+            is_new_style_constructor(),
+            extra...);
+    }
+};
+
+// Implementing class for py::init_alias<...>()
+template <typename... Args>
+struct alias_constructor {
+    template <
+        typename Class,
+        typename... Extra,
+        enable_if_t<Class::has_alias && std::is_constructible<Alias<Class>, Args...>::value, int>
+        = 0>
+    static void execute(Class &cl, const Extra &...extra) {
+        cl.def(
+            "__init__",
+            [](value_and_holder &v_h, Args... args) {
+                v_h.value_ptr()
+                    = construct_or_initialize<Alias<Class>>(std::forward<Args>(args)...);
+            },
+            is_new_style_constructor(),
+            extra...);
+    }
+};
+
+// Implementation class for py::init(Func) and py::init(Func, AliasFunc)
+template <typename CFunc,
+          typename AFunc = void_type (*)(),
+          typename = function_signature_t<CFunc>,
+          typename = function_signature_t<AFunc>>
+struct factory;
+
+// Specialization for py::init(Func)
+template <typename Func, typename Return, typename... Args>
+struct factory<Func, void_type (*)(), Return(Args...)> {
+    remove_reference_t<Func> class_factory;
+
+    // NOLINTNEXTLINE(google-explicit-constructor)
+    factory(Func &&f) : class_factory(std::forward<Func>(f)) {}
+
+    // The given class either has no alias or has no separate alias factory;
+    // this always constructs the class itself.  If the class is registered with an alias
+    // type and an alias instance is needed (i.e. because the final type is a Python class
+    // inheriting from the C++ type) the returned value needs to either already be an alias
+    // instance, or the alias needs to be constructible from a `Class &&` argument.
+    template <typename Class, typename... Extra>
+    void execute(Class &cl, const Extra &...extra) && {
+#if defined(PYBIND11_CPP14)
+        cl.def(
+            "__init__",
+            [func = std::move(class_factory)]
+#else
+        auto &func = class_factory;
+        cl.def(
+            "__init__",
+            [func]
+#endif
+            (value_and_holder &v_h, Args... args) {
+                construct<Class>(
+                    v_h, func(std::forward<Args>(args)...), Py_TYPE(v_h.inst) != v_h.type->type);
+            },
+            is_new_style_constructor(),
+            extra...);
+    }
+};
+
+// Specialization for py::init(Func, AliasFunc)
+template <typename CFunc,
+          typename AFunc,
+          typename CReturn,
+          typename... CArgs,
+          typename AReturn,
+          typename... AArgs>
+struct factory<CFunc, AFunc, CReturn(CArgs...), AReturn(AArgs...)> {
+    static_assert(sizeof...(CArgs) == sizeof...(AArgs),
+                  "pybind11::init(class_factory, alias_factory): class and alias factories "
+                  "must have identical argument signatures");
+    static_assert(all_of<std::is_same<CArgs, AArgs>...>::value,
+                  "pybind11::init(class_factory, alias_factory): class and alias factories "
+                  "must have identical argument signatures");
+
+    remove_reference_t<CFunc> class_factory;
+    remove_reference_t<AFunc> alias_factory;
+
+    factory(CFunc &&c, AFunc &&a)
+        : class_factory(std::forward<CFunc>(c)), alias_factory(std::forward<AFunc>(a)) {}
+
+    // The class factory is called when the `self` type passed to `__init__` is the direct
+    // class (i.e. not inherited), the alias factory when `self` is a Python-side subtype.
+    template <typename Class, typename... Extra>
+    void execute(Class &cl, const Extra &...extra) && {
+        static_assert(Class::has_alias,
+                      "The two-argument version of `py::init()` can "
+                      "only be used if the class has an alias");
+#if defined(PYBIND11_CPP14)
+        cl.def(
+            "__init__",
+            [class_func = std::move(class_factory), alias_func = std::move(alias_factory)]
+#else
+        auto &class_func = class_factory;
+        auto &alias_func = alias_factory;
+        cl.def(
+            "__init__",
+            [class_func, alias_func]
+#endif
+            (value_and_holder &v_h, CArgs... args) {
+                if (Py_TYPE(v_h.inst) == v_h.type->type) {
+                    // If the instance type equals the registered type we don't have inheritance,
+                    // so don't need the alias and can construct using the class function:
+                    construct<Class>(v_h, class_func(std::forward<CArgs>(args)...), false);
+                } else {
+                    construct<Class>(v_h, alias_func(std::forward<CArgs>(args)...), true);
+                }
+            },
+            is_new_style_constructor(),
+            extra...);
+    }
+};
+
+/// Set just the C++ state. Same as `__init__`.
+template <typename Class, typename T>
+void setstate(value_and_holder &v_h, T &&result, bool need_alias) {
+    construct<Class>(v_h, std::forward<T>(result), need_alias);
+}
+
+/// Set both the C++ and Python states
+template <typename Class,
+          typename T,
+          typename O,
+          enable_if_t<std::is_convertible<O, handle>::value, int> = 0>
+void setstate(value_and_holder &v_h, std::pair<T, O> &&result, bool need_alias) {
+    construct<Class>(v_h, std::move(result.first), need_alias);
+    auto d = handle(result.second);
+    if (PyDict_Check(d.ptr()) && PyDict_Size(d.ptr()) == 0) {
+        // Skipping setattr below, to not force use of py::dynamic_attr() for Class unnecessarily.
+        // See PR #2972 for details.
+        return;
+    }
+    setattr((PyObject *) v_h.inst, "__dict__", d);
+}
+
+/// Implementation for py::pickle(GetState, SetState)
+template <typename Get,
+          typename Set,
+          typename = function_signature_t<Get>,
+          typename = function_signature_t<Set>>
+struct pickle_factory;
+
+template <typename Get,
+          typename Set,
+          typename RetState,
+          typename Self,
+          typename NewInstance,
+          typename ArgState>
+struct pickle_factory<Get, Set, RetState(Self), NewInstance(ArgState)> {
+    static_assert(std::is_same<intrinsic_t<RetState>, intrinsic_t<ArgState>>::value,
+                  "The type returned by `__getstate__` must be the same "
+                  "as the argument accepted by `__setstate__`");
+
+    remove_reference_t<Get> get;
+    remove_reference_t<Set> set;
+
+    pickle_factory(Get get, Set set) : get(std::forward<Get>(get)), set(std::forward<Set>(set)) {}
+
+    template <typename Class, typename... Extra>
+    void execute(Class &cl, const Extra &...extra) && {
+        cl.def("__getstate__", std::move(get));
+
+#if defined(PYBIND11_CPP14)
+        cl.def(
+            "__setstate__",
+            [func = std::move(set)]
+#else
+        auto &func = set;
+        cl.def(
+            "__setstate__",
+            [func]
+#endif
+            (value_and_holder &v_h, ArgState state) {
+                setstate<Class>(
+                    v_h, func(std::forward<ArgState>(state)), Py_TYPE(v_h.inst) != v_h.type->type);
+            },
+            is_new_style_constructor(),
+            extra...);
+    }
+};
+
+PYBIND11_NAMESPACE_END(initimpl)
+PYBIND11_NAMESPACE_END(detail)
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/detail/internals.h

@@ -0,0 +1,766 @@
+/*
+    pybind11/detail/internals.h: Internal data structure and related functions
+
+    Copyright (c) 2017 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "common.h"
+
+#if defined(PYBIND11_SIMPLE_GIL_MANAGEMENT)
+#    include <pybind11/gil.h>
+#endif
+
+#include <pybind11/pytypes.h>
+
+#include <exception>
+#include <mutex>
+#include <thread>
+
+/// Tracks the `internals` and `type_info` ABI version independent of the main library version.
+///
+/// Some portions of the code use an ABI that is conditional depending on this
+/// version number.  That allows ABI-breaking changes to be "pre-implemented".
+/// Once the default version number is incremented, the conditional logic that
+/// no longer applies can be removed.  Additionally, users that need not
+/// maintain ABI compatibility can increase the version number in order to take
+/// advantage of any functionality/efficiency improvements that depend on the
+/// newer ABI.
+///
+/// WARNING: If you choose to manually increase the ABI version, note that
+/// pybind11 may not be tested as thoroughly with a non-default ABI version, and
+/// further ABI-incompatible changes may be made before the ABI is officially
+/// changed to the new version.
+#ifndef PYBIND11_INTERNALS_VERSION
+#    if PY_VERSION_HEX >= 0x030C0000 || defined(_MSC_VER)
+// Version bump for Python 3.12+, before first 3.12 beta release.
+// Version bump for MSVC piggy-backed on PR #4779. See comments there.
+#        define PYBIND11_INTERNALS_VERSION 5
+#    else
+#        define PYBIND11_INTERNALS_VERSION 4
+#    endif
+#endif
+
+// This requirement is mainly to reduce the support burden (see PR #4570).
+static_assert(PY_VERSION_HEX < 0x030C0000 || PYBIND11_INTERNALS_VERSION >= 5,
+              "pybind11 ABI version 5 is the minimum for Python 3.12+");
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+using ExceptionTranslator = void (*)(std::exception_ptr);
+
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+constexpr const char *internals_function_record_capsule_name = "pybind11_function_record_capsule";
+
+// Forward declarations
+inline PyTypeObject *make_static_property_type();
+inline PyTypeObject *make_default_metaclass();
+inline PyObject *make_object_base_type(PyTypeObject *metaclass);
+
+// The old Python Thread Local Storage (TLS) API is deprecated in Python 3.7 in favor of the new
+// Thread Specific Storage (TSS) API.
+// Avoid unnecessary allocation of `Py_tss_t`, since we cannot use
+// `Py_LIMITED_API` anyway.
+#if PYBIND11_INTERNALS_VERSION > 4
+#    define PYBIND11_TLS_KEY_REF Py_tss_t &
+#    if defined(__clang__)
+#        define PYBIND11_TLS_KEY_INIT(var)                                                        \
+            _Pragma("clang diagnostic push")                                         /**/         \
+                _Pragma("clang diagnostic ignored \"-Wmissing-field-initializers\"") /**/         \
+                Py_tss_t var                                                                      \
+                = Py_tss_NEEDS_INIT;                                                              \
+            _Pragma("clang diagnostic pop")
+#    elif defined(__GNUC__) && !defined(__INTEL_COMPILER)
+#        define PYBIND11_TLS_KEY_INIT(var)                                                        \
+            _Pragma("GCC diagnostic push")                                         /**/           \
+                _Pragma("GCC diagnostic ignored \"-Wmissing-field-initializers\"") /**/           \
+                Py_tss_t var                                                                      \
+                = Py_tss_NEEDS_INIT;                                                              \
+            _Pragma("GCC diagnostic pop")
+#    else
+#        define PYBIND11_TLS_KEY_INIT(var) Py_tss_t var = Py_tss_NEEDS_INIT;
+#    endif
+#    define PYBIND11_TLS_KEY_CREATE(var) (PyThread_tss_create(&(var)) == 0)
+#    define PYBIND11_TLS_GET_VALUE(key) PyThread_tss_get(&(key))
+#    define PYBIND11_TLS_REPLACE_VALUE(key, value) PyThread_tss_set(&(key), (value))
+#    define PYBIND11_TLS_DELETE_VALUE(key) PyThread_tss_set(&(key), nullptr)
+#    define PYBIND11_TLS_FREE(key) PyThread_tss_delete(&(key))
+#else
+#    define PYBIND11_TLS_KEY_REF Py_tss_t *
+#    define PYBIND11_TLS_KEY_INIT(var) Py_tss_t *var = nullptr;
+#    define PYBIND11_TLS_KEY_CREATE(var)                                                          \
+        (((var) = PyThread_tss_alloc()) != nullptr && (PyThread_tss_create((var)) == 0))
+#    define PYBIND11_TLS_GET_VALUE(key) PyThread_tss_get((key))
+#    define PYBIND11_TLS_REPLACE_VALUE(key, value) PyThread_tss_set((key), (value))
+#    define PYBIND11_TLS_DELETE_VALUE(key) PyThread_tss_set((key), nullptr)
+#    define PYBIND11_TLS_FREE(key) PyThread_tss_free(key)
+#endif
+
+// Python loads modules by default with dlopen with the RTLD_LOCAL flag; under libc++ and possibly
+// other STLs, this means `typeid(A)` from one module won't equal `typeid(A)` from another module
+// even when `A` is the same, non-hidden-visibility type (e.g. from a common include).  Under
+// libstdc++, this doesn't happen: equality and the type_index hash are based on the type name,
+// which works.  If not under a known-good stl, provide our own name-based hash and equality
+// functions that use the type name.
+#if (PYBIND11_INTERNALS_VERSION <= 4 && defined(__GLIBCXX__))                                     \
+    || (PYBIND11_INTERNALS_VERSION >= 5 && !defined(_LIBCPP_VERSION))
+inline bool same_type(const std::type_info &lhs, const std::type_info &rhs) { return lhs == rhs; }
+using type_hash = std::hash<std::type_index>;
+using type_equal_to = std::equal_to<std::type_index>;
+#else
+inline bool same_type(const std::type_info &lhs, const std::type_info &rhs) {
+    return lhs.name() == rhs.name() || std::strcmp(lhs.name(), rhs.name()) == 0;
+}
+
+struct type_hash {
+    size_t operator()(const std::type_index &t) const {
+        size_t hash = 5381;
+        const char *ptr = t.name();
+        while (auto c = static_cast<unsigned char>(*ptr++)) {
+            hash = (hash * 33) ^ c;
+        }
+        return hash;
+    }
+};
+
+struct type_equal_to {
+    bool operator()(const std::type_index &lhs, const std::type_index &rhs) const {
+        return lhs.name() == rhs.name() || std::strcmp(lhs.name(), rhs.name()) == 0;
+    }
+};
+#endif
+
+template <typename value_type>
+using type_map = std::unordered_map<std::type_index, value_type, type_hash, type_equal_to>;
+
+struct override_hash {
+    inline size_t operator()(const std::pair<const PyObject *, const char *> &v) const {
+        size_t value = std::hash<const void *>()(v.first);
+        value ^= std::hash<const void *>()(v.second) + 0x9e3779b9 + (value << 6) + (value >> 2);
+        return value;
+    }
+};
+
+using instance_map = std::unordered_multimap<const void *, instance *>;
+
+#ifdef Py_GIL_DISABLED
+// Wrapper around PyMutex to provide BasicLockable semantics
+class pymutex {
+    PyMutex mutex;
+
+public:
+    pymutex() : mutex({}) {}
+    void lock() { PyMutex_Lock(&mutex); }
+    void unlock() { PyMutex_Unlock(&mutex); }
+};
+
+// Instance map shards are used to reduce mutex contention in free-threaded Python.
+struct instance_map_shard {
+    instance_map registered_instances;
+    pymutex mutex;
+    // alignas(64) would be better, but causes compile errors in macOS before 10.14 (see #5200)
+    char padding[64 - (sizeof(instance_map) + sizeof(pymutex)) % 64];
+};
+
+static_assert(sizeof(instance_map_shard) % 64 == 0,
+              "instance_map_shard size is not a multiple of 64 bytes");
+#endif
+
+/// Internal data structure used to track registered instances and types.
+/// Whenever binary incompatible changes are made to this structure,
+/// `PYBIND11_INTERNALS_VERSION` must be incremented.
+struct internals {
+#ifdef Py_GIL_DISABLED
+    pymutex mutex;
+#endif
+    // std::type_index -> pybind11's type information
+    type_map<type_info *> registered_types_cpp;
+    // PyTypeObject* -> base type_info(s)
+    std::unordered_map<PyTypeObject *, std::vector<type_info *>> registered_types_py;
+#ifdef Py_GIL_DISABLED
+    std::unique_ptr<instance_map_shard[]> instance_shards; // void * -> instance*
+    size_t instance_shards_mask;
+#else
+    instance_map registered_instances; // void * -> instance*
+#endif
+    std::unordered_set<std::pair<const PyObject *, const char *>, override_hash>
+        inactive_override_cache;
+    type_map<std::vector<bool (*)(PyObject *, void *&)>> direct_conversions;
+    std::unordered_map<const PyObject *, std::vector<PyObject *>> patients;
+    std::forward_list<ExceptionTranslator> registered_exception_translators;
+    std::unordered_map<std::string, void *> shared_data; // Custom data to be shared across
+                                                         // extensions
+#if PYBIND11_INTERNALS_VERSION == 4
+    std::vector<PyObject *> unused_loader_patient_stack_remove_at_v5;
+#endif
+    std::forward_list<std::string> static_strings; // Stores the std::strings backing
+                                                   // detail::c_str()
+    PyTypeObject *static_property_type;
+    PyTypeObject *default_metaclass;
+    PyObject *instance_base;
+    // Unused if PYBIND11_SIMPLE_GIL_MANAGEMENT is defined:
+    PYBIND11_TLS_KEY_INIT(tstate)
+#if PYBIND11_INTERNALS_VERSION > 4
+    PYBIND11_TLS_KEY_INIT(loader_life_support_tls_key)
+#endif // PYBIND11_INTERNALS_VERSION > 4
+    // Unused if PYBIND11_SIMPLE_GIL_MANAGEMENT is defined:
+    PyInterpreterState *istate = nullptr;
+
+#if PYBIND11_INTERNALS_VERSION > 4
+    // Note that we have to use a std::string to allocate memory to ensure a unique address
+    // We want unique addresses since we use pointer equality to compare function records
+    std::string function_record_capsule_name = internals_function_record_capsule_name;
+#endif
+
+    internals() = default;
+    internals(const internals &other) = delete;
+    internals &operator=(const internals &other) = delete;
+    ~internals() {
+#if PYBIND11_INTERNALS_VERSION > 4
+        PYBIND11_TLS_FREE(loader_life_support_tls_key);
+#endif // PYBIND11_INTERNALS_VERSION > 4
+
+        // This destructor is called *after* Py_Finalize() in finalize_interpreter().
+        // That *SHOULD BE* fine. The following details what happens when PyThread_tss_free is
+        // called. PYBIND11_TLS_FREE is PyThread_tss_free on python 3.7+. On older python, it does
+        // nothing. PyThread_tss_free calls PyThread_tss_delete and PyMem_RawFree.
+        // PyThread_tss_delete just calls TlsFree (on Windows) or pthread_key_delete (on *NIX).
+        // Neither of those have anything to do with CPython internals. PyMem_RawFree *requires*
+        // that the `tstate` be allocated with the CPython allocator.
+        PYBIND11_TLS_FREE(tstate);
+    }
+};
+
+/// Additional type information which does not fit into the PyTypeObject.
+/// Changes to this struct also require bumping `PYBIND11_INTERNALS_VERSION`.
+struct type_info {
+    PyTypeObject *type;
+    const std::type_info *cpptype;
+    size_t type_size, type_align, holder_size_in_ptrs;
+    void *(*operator_new)(size_t);
+    void (*init_instance)(instance *, const void *);
+    void (*dealloc)(value_and_holder &v_h);
+    std::vector<PyObject *(*) (PyObject *, PyTypeObject *)> implicit_conversions;
+    std::vector<std::pair<const std::type_info *, void *(*) (void *)>> implicit_casts;
+    std::vector<bool (*)(PyObject *, void *&)> *direct_conversions;
+    buffer_info *(*get_buffer)(PyObject *, void *) = nullptr;
+    void *get_buffer_data = nullptr;
+    void *(*module_local_load)(PyObject *, const type_info *) = nullptr;
+    /* A simple type never occurs as a (direct or indirect) parent
+     * of a class that makes use of multiple inheritance.
+     * A type can be simple even if it has non-simple ancestors as long as it has no descendants.
+     */
+    bool simple_type : 1;
+    /* True if there is no multiple inheritance in this type's inheritance tree */
+    bool simple_ancestors : 1;
+    /* for base vs derived holder_type checks */
+    bool default_holder : 1;
+    /* true if this is a type registered with py::module_local */
+    bool module_local : 1;
+};
+
+/// On MSVC, debug and release builds are not ABI-compatible!
+#if defined(_MSC_VER) && defined(_DEBUG)
+#    define PYBIND11_BUILD_TYPE "_debug"
+#else
+#    define PYBIND11_BUILD_TYPE ""
+#endif
+
+/// Let's assume that different compilers are ABI-incompatible.
+/// A user can manually set this string if they know their
+/// compiler is compatible.
+#ifndef PYBIND11_COMPILER_TYPE
+#    if defined(_MSC_VER)
+#        define PYBIND11_COMPILER_TYPE "_msvc"
+#    elif defined(__INTEL_COMPILER)
+#        define PYBIND11_COMPILER_TYPE "_icc"
+#    elif defined(__clang__)
+#        define PYBIND11_COMPILER_TYPE "_clang"
+#    elif defined(__PGI)
+#        define PYBIND11_COMPILER_TYPE "_pgi"
+#    elif defined(__MINGW32__)
+#        define PYBIND11_COMPILER_TYPE "_mingw"
+#    elif defined(__CYGWIN__)
+#        define PYBIND11_COMPILER_TYPE "_gcc_cygwin"
+#    elif defined(__GNUC__)
+#        define PYBIND11_COMPILER_TYPE "_gcc"
+#    else
+#        define PYBIND11_COMPILER_TYPE "_unknown"
+#    endif
+#endif
+
+/// Also standard libs
+#ifndef PYBIND11_STDLIB
+#    if defined(_LIBCPP_VERSION)
+#        define PYBIND11_STDLIB "_libcpp"
+#    elif defined(__GLIBCXX__) || defined(__GLIBCPP__)
+#        define PYBIND11_STDLIB "_libstdcpp"
+#    else
+#        define PYBIND11_STDLIB ""
+#    endif
+#endif
+
+/// On Linux/OSX, changes in __GXX_ABI_VERSION__ indicate ABI incompatibility.
+/// On MSVC, changes in _MSC_VER may indicate ABI incompatibility (#2898).
+#ifndef PYBIND11_BUILD_ABI
+#    if defined(__GXX_ABI_VERSION)
+#        define PYBIND11_BUILD_ABI "_cxxabi" PYBIND11_TOSTRING(__GXX_ABI_VERSION)
+#    elif defined(_MSC_VER)
+#        define PYBIND11_BUILD_ABI "_mscver" PYBIND11_TOSTRING(_MSC_VER)
+#    else
+#        define PYBIND11_BUILD_ABI ""
+#    endif
+#endif
+
+#ifndef PYBIND11_INTERNALS_KIND
+#    define PYBIND11_INTERNALS_KIND ""
+#endif
+
+#define PYBIND11_PLATFORM_ABI_ID                                                                  \
+    PYBIND11_INTERNALS_KIND PYBIND11_COMPILER_TYPE PYBIND11_STDLIB PYBIND11_BUILD_ABI             \
+        PYBIND11_BUILD_TYPE
+
+#define PYBIND11_INTERNALS_ID                                                                     \
+    "__pybind11_internals_v" PYBIND11_TOSTRING(PYBIND11_INTERNALS_VERSION)                        \
+        PYBIND11_PLATFORM_ABI_ID "__"
+
+#define PYBIND11_MODULE_LOCAL_ID                                                                  \
+    "__pybind11_module_local_v" PYBIND11_TOSTRING(PYBIND11_INTERNALS_VERSION)                     \
+        PYBIND11_PLATFORM_ABI_ID "__"
+
+/// Each module locally stores a pointer to the `internals` data. The data
+/// itself is shared among modules with the same `PYBIND11_INTERNALS_ID`.
+inline internals **&get_internals_pp() {
+    static internals **internals_pp = nullptr;
+    return internals_pp;
+}
+
+// forward decl
+inline void translate_exception(std::exception_ptr);
+
+template <class T,
+          enable_if_t<std::is_same<std::nested_exception, remove_cvref_t<T>>::value, int> = 0>
+bool handle_nested_exception(const T &exc, const std::exception_ptr &p) {
+    std::exception_ptr nested = exc.nested_ptr();
+    if (nested != nullptr && nested != p) {
+        translate_exception(nested);
+        return true;
+    }
+    return false;
+}
+
+template <class T,
+          enable_if_t<!std::is_same<std::nested_exception, remove_cvref_t<T>>::value, int> = 0>
+bool handle_nested_exception(const T &exc, const std::exception_ptr &p) {
+    if (const auto *nep = dynamic_cast<const std::nested_exception *>(std::addressof(exc))) {
+        return handle_nested_exception(*nep, p);
+    }
+    return false;
+}
+
+inline bool raise_err(PyObject *exc_type, const char *msg) {
+    if (PyErr_Occurred()) {
+        raise_from(exc_type, msg);
+        return true;
+    }
+    set_error(exc_type, msg);
+    return false;
+}
+
+inline void translate_exception(std::exception_ptr p) {
+    if (!p) {
+        return;
+    }
+    try {
+        std::rethrow_exception(p);
+    } catch (error_already_set &e) {
+        handle_nested_exception(e, p);
+        e.restore();
+        return;
+    } catch (const builtin_exception &e) {
+        // Could not use template since it's an abstract class.
+        if (const auto *nep = dynamic_cast<const std::nested_exception *>(std::addressof(e))) {
+            handle_nested_exception(*nep, p);
+        }
+        e.set_error();
+        return;
+    } catch (const std::bad_alloc &e) {
+        handle_nested_exception(e, p);
+        raise_err(PyExc_MemoryError, e.what());
+        return;
+    } catch (const std::domain_error &e) {
+        handle_nested_exception(e, p);
+        raise_err(PyExc_ValueError, e.what());
+        return;
+    } catch (const std::invalid_argument &e) {
+        handle_nested_exception(e, p);
+        raise_err(PyExc_ValueError, e.what());
+        return;
+    } catch (const std::length_error &e) {
+        handle_nested_exception(e, p);
+        raise_err(PyExc_ValueError, e.what());
+        return;
+    } catch (const std::out_of_range &e) {
+        handle_nested_exception(e, p);
+        raise_err(PyExc_IndexError, e.what());
+        return;
+    } catch (const std::range_error &e) {
+        handle_nested_exception(e, p);
+        raise_err(PyExc_ValueError, e.what());
+        return;
+    } catch (const std::overflow_error &e) {
+        handle_nested_exception(e, p);
+        raise_err(PyExc_OverflowError, e.what());
+        return;
+    } catch (const std::exception &e) {
+        handle_nested_exception(e, p);
+        raise_err(PyExc_RuntimeError, e.what());
+        return;
+    } catch (const std::nested_exception &e) {
+        handle_nested_exception(e, p);
+        raise_err(PyExc_RuntimeError, "Caught an unknown nested exception!");
+        return;
+    } catch (...) {
+        raise_err(PyExc_RuntimeError, "Caught an unknown exception!");
+        return;
+    }
+}
+
+#if !defined(__GLIBCXX__)
+inline void translate_local_exception(std::exception_ptr p) {
+    try {
+        if (p) {
+            std::rethrow_exception(p);
+        }
+    } catch (error_already_set &e) {
+        e.restore();
+        return;
+    } catch (const builtin_exception &e) {
+        e.set_error();
+        return;
+    }
+}
+#endif
+
+inline object get_python_state_dict() {
+    object state_dict;
+#if PYBIND11_INTERNALS_VERSION <= 4 || PY_VERSION_HEX < 0x03080000 || defined(PYPY_VERSION)
+    state_dict = reinterpret_borrow<object>(PyEval_GetBuiltins());
+#else
+#    if PY_VERSION_HEX < 0x03090000
+    PyInterpreterState *istate = _PyInterpreterState_Get();
+#    else
+    PyInterpreterState *istate = PyInterpreterState_Get();
+#    endif
+    if (istate) {
+        state_dict = reinterpret_borrow<object>(PyInterpreterState_GetDict(istate));
+    }
+#endif
+    if (!state_dict) {
+        raise_from(PyExc_SystemError, "pybind11::detail::get_python_state_dict() FAILED");
+        throw error_already_set();
+    }
+    return state_dict;
+}
+
+inline object get_internals_obj_from_state_dict(handle state_dict) {
+    return reinterpret_steal<object>(
+        dict_getitemstringref(state_dict.ptr(), PYBIND11_INTERNALS_ID));
+}
+
+inline internals **get_internals_pp_from_capsule(handle obj) {
+    void *raw_ptr = PyCapsule_GetPointer(obj.ptr(), /*name=*/nullptr);
+    if (raw_ptr == nullptr) {
+        raise_from(PyExc_SystemError, "pybind11::detail::get_internals_pp_from_capsule() FAILED");
+        throw error_already_set();
+    }
+    return static_cast<internals **>(raw_ptr);
+}
+
+inline uint64_t round_up_to_next_pow2(uint64_t x) {
+    // Round-up to the next power of two.
+    // See https://graphics.stanford.edu/~seander/bithacks.html#RoundUpPowerOf2
+    x--;
+    x |= (x >> 1);
+    x |= (x >> 2);
+    x |= (x >> 4);
+    x |= (x >> 8);
+    x |= (x >> 16);
+    x |= (x >> 32);
+    x++;
+    return x;
+}
+
+/// Return a reference to the current `internals` data
+PYBIND11_NOINLINE internals &get_internals() {
+    auto **&internals_pp = get_internals_pp();
+    if (internals_pp && *internals_pp) {
+        return **internals_pp;
+    }
+
+#if defined(PYBIND11_SIMPLE_GIL_MANAGEMENT)
+    gil_scoped_acquire gil;
+#else
+    // Ensure that the GIL is held since we will need to make Python calls.
+    // Cannot use py::gil_scoped_acquire here since that constructor calls get_internals.
+    struct gil_scoped_acquire_local {
+        gil_scoped_acquire_local() : state(PyGILState_Ensure()) {}
+        gil_scoped_acquire_local(const gil_scoped_acquire_local &) = delete;
+        gil_scoped_acquire_local &operator=(const gil_scoped_acquire_local &) = delete;
+        ~gil_scoped_acquire_local() { PyGILState_Release(state); }
+        const PyGILState_STATE state;
+    } gil;
+#endif
+    error_scope err_scope;
+
+    dict state_dict = get_python_state_dict();
+    if (object internals_obj = get_internals_obj_from_state_dict(state_dict)) {
+        internals_pp = get_internals_pp_from_capsule(internals_obj);
+    }
+    if (internals_pp && *internals_pp) {
+        // We loaded the internals through `state_dict`, which means that our `error_already_set`
+        // and `builtin_exception` may be different local classes than the ones set up in the
+        // initial exception translator, below, so add another for our local exception classes.
+        //
+        // libstdc++ doesn't require this (types there are identified only by name)
+        // libc++ with CPython doesn't require this (types are explicitly exported)
+        // libc++ with PyPy still need it, awaiting further investigation
+#if !defined(__GLIBCXX__)
+        (*internals_pp)->registered_exception_translators.push_front(&translate_local_exception);
+#endif
+    } else {
+        if (!internals_pp) {
+            internals_pp = new internals *();
+        }
+        auto *&internals_ptr = *internals_pp;
+        internals_ptr = new internals();
+
+        PyThreadState *tstate = PyThreadState_Get();
+        // NOLINTNEXTLINE(bugprone-assignment-in-if-condition)
+        if (!PYBIND11_TLS_KEY_CREATE(internals_ptr->tstate)) {
+            pybind11_fail("get_internals: could not successfully initialize the tstate TSS key!");
+        }
+        PYBIND11_TLS_REPLACE_VALUE(internals_ptr->tstate, tstate);
+
+#if PYBIND11_INTERNALS_VERSION > 4
+        // NOLINTNEXTLINE(bugprone-assignment-in-if-condition)
+        if (!PYBIND11_TLS_KEY_CREATE(internals_ptr->loader_life_support_tls_key)) {
+            pybind11_fail("get_internals: could not successfully initialize the "
+                          "loader_life_support TSS key!");
+        }
+#endif
+        internals_ptr->istate = tstate->interp;
+        state_dict[PYBIND11_INTERNALS_ID] = capsule(reinterpret_cast<void *>(internals_pp));
+        internals_ptr->registered_exception_translators.push_front(&translate_exception);
+        internals_ptr->static_property_type = make_static_property_type();
+        internals_ptr->default_metaclass = make_default_metaclass();
+        internals_ptr->instance_base = make_object_base_type(internals_ptr->default_metaclass);
+#ifdef Py_GIL_DISABLED
+        // Scale proportional to the number of cores. 2x is a heuristic to reduce contention.
+        auto num_shards
+            = static_cast<size_t>(round_up_to_next_pow2(2 * std::thread::hardware_concurrency()));
+        if (num_shards == 0) {
+            num_shards = 1;
+        }
+        internals_ptr->instance_shards.reset(new instance_map_shard[num_shards]);
+        internals_ptr->instance_shards_mask = num_shards - 1;
+#endif // Py_GIL_DISABLED
+    }
+    return **internals_pp;
+}
+
+// the internals struct (above) is shared between all the modules. local_internals are only
+// for a single module. Any changes made to internals may require an update to
+// PYBIND11_INTERNALS_VERSION, breaking backwards compatibility. local_internals is, by design,
+// restricted to a single module. Whether a module has local internals or not should not
+// impact any other modules, because the only things accessing the local internals is the
+// module that contains them.
+struct local_internals {
+    type_map<type_info *> registered_types_cpp;
+    std::forward_list<ExceptionTranslator> registered_exception_translators;
+#if PYBIND11_INTERNALS_VERSION == 4
+
+    // For ABI compatibility, we can't store the loader_life_support TLS key in
+    // the `internals` struct directly.  Instead, we store it in `shared_data` and
+    // cache a copy in `local_internals`.  If we allocated a separate TLS key for
+    // each instance of `local_internals`, we could end up allocating hundreds of
+    // TLS keys if hundreds of different pybind11 modules are loaded (which is a
+    // plausible number).
+    PYBIND11_TLS_KEY_INIT(loader_life_support_tls_key)
+
+    // Holds the shared TLS key for the loader_life_support stack.
+    struct shared_loader_life_support_data {
+        PYBIND11_TLS_KEY_INIT(loader_life_support_tls_key)
+        shared_loader_life_support_data() {
+            // NOLINTNEXTLINE(bugprone-assignment-in-if-condition)
+            if (!PYBIND11_TLS_KEY_CREATE(loader_life_support_tls_key)) {
+                pybind11_fail("local_internals: could not successfully initialize the "
+                              "loader_life_support TLS key!");
+            }
+        }
+        // We can't help but leak the TLS key, because Python never unloads extension modules.
+    };
+
+    local_internals() {
+        auto &internals = get_internals();
+        // Get or create the `loader_life_support_stack_key`.
+        auto &ptr = internals.shared_data["_life_support"];
+        if (!ptr) {
+            ptr = new shared_loader_life_support_data;
+        }
+        loader_life_support_tls_key
+            = static_cast<shared_loader_life_support_data *>(ptr)->loader_life_support_tls_key;
+    }
+#endif //  PYBIND11_INTERNALS_VERSION == 4
+};
+
+/// Works like `get_internals`, but for things which are locally registered.
+inline local_internals &get_local_internals() {
+    // Current static can be created in the interpreter finalization routine. If the later will be
+    // destroyed in another static variable destructor, creation of this static there will cause
+    // static deinitialization fiasco. In order to avoid it we avoid destruction of the
+    // local_internals static. One can read more about the problem and current solution here:
+    // https://google.github.io/styleguide/cppguide.html#Static_and_Global_Variables
+    static auto *locals = new local_internals();
+    return *locals;
+}
+
+#ifdef Py_GIL_DISABLED
+#    define PYBIND11_LOCK_INTERNALS(internals) std::unique_lock<pymutex> lock((internals).mutex)
+#else
+#    define PYBIND11_LOCK_INTERNALS(internals)
+#endif
+
+template <typename F>
+inline auto with_internals(const F &cb) -> decltype(cb(get_internals())) {
+    auto &internals = get_internals();
+    PYBIND11_LOCK_INTERNALS(internals);
+    return cb(internals);
+}
+
+inline std::uint64_t mix64(std::uint64_t z) {
+    // David Stafford's variant 13 of the MurmurHash3 finalizer popularized
+    // by the SplitMix PRNG.
+    // https://zimbry.blogspot.com/2011/09/better-bit-mixing-improving-on.html
+    z = (z ^ (z >> 30)) * 0xbf58476d1ce4e5b9;
+    z = (z ^ (z >> 27)) * 0x94d049bb133111eb;
+    return z ^ (z >> 31);
+}
+
+template <typename F>
+inline auto with_instance_map(const void *ptr,
+                              const F &cb) -> decltype(cb(std::declval<instance_map &>())) {
+    auto &internals = get_internals();
+
+#ifdef Py_GIL_DISABLED
+    // Hash address to compute shard, but ignore low bits. We'd like allocations
+    // from the same thread/core to map to the same shard and allocations from
+    // other threads/cores to map to other shards. Using the high bits is a good
+    // heuristic because memory allocators often have a per-thread
+    // arena/superblock/segment from which smaller allocations are served.
+    auto addr = reinterpret_cast<std::uintptr_t>(ptr);
+    auto hash = mix64(static_cast<std::uint64_t>(addr >> 20));
+    auto idx = static_cast<size_t>(hash & internals.instance_shards_mask);
+
+    auto &shard = internals.instance_shards[idx];
+    std::unique_lock<pymutex> lock(shard.mutex);
+    return cb(shard.registered_instances);
+#else
+    (void) ptr;
+    return cb(internals.registered_instances);
+#endif
+}
+
+// Returns the number of registered instances for testing purposes.  The result may not be
+// consistent if other threads are registering or unregistering instances concurrently.
+inline size_t num_registered_instances() {
+    auto &internals = get_internals();
+#ifdef Py_GIL_DISABLED
+    size_t count = 0;
+    for (size_t i = 0; i <= internals.instance_shards_mask; ++i) {
+        auto &shard = internals.instance_shards[i];
+        std::unique_lock<pymutex> lock(shard.mutex);
+        count += shard.registered_instances.size();
+    }
+    return count;
+#else
+    return internals.registered_instances.size();
+#endif
+}
+
+/// Constructs a std::string with the given arguments, stores it in `internals`, and returns its
+/// `c_str()`.  Such strings objects have a long storage duration -- the internal strings are only
+/// cleared when the program exits or after interpreter shutdown (when embedding), and so are
+/// suitable for c-style strings needed by Python internals (such as PyTypeObject's tp_name).
+template <typename... Args>
+const char *c_str(Args &&...args) {
+    // GCC 4.8 doesn't like parameter unpack within lambda capture, so use
+    // PYBIND11_LOCK_INTERNALS.
+    auto &internals = get_internals();
+    PYBIND11_LOCK_INTERNALS(internals);
+    auto &strings = internals.static_strings;
+    strings.emplace_front(std::forward<Args>(args)...);
+    return strings.front().c_str();
+}
+
+inline const char *get_function_record_capsule_name() {
+#if PYBIND11_INTERNALS_VERSION > 4
+    return get_internals().function_record_capsule_name.c_str();
+#else
+    return nullptr;
+#endif
+}
+
+// Determine whether or not the following capsule contains a pybind11 function record.
+// Note that we use `internals` to make sure that only ABI compatible records are touched.
+//
+// This check is currently used in two places:
+// - An important optimization in functional.h to avoid overhead in C++ -> Python -> C++
+// - The sibling feature of cpp_function to allow overloads
+inline bool is_function_record_capsule(const capsule &cap) {
+    // Pointer equality as we rely on internals() to ensure unique pointers
+    return cap.name() == get_function_record_capsule_name();
+}
+
+PYBIND11_NAMESPACE_END(detail)
+
+/// Returns a named pointer that is shared among all extension modules (using the same
+/// pybind11 version) running in the current interpreter. Names starting with underscores
+/// are reserved for internal usage. Returns `nullptr` if no matching entry was found.
+PYBIND11_NOINLINE void *get_shared_data(const std::string &name) {
+    return detail::with_internals([&](detail::internals &internals) {
+        auto it = internals.shared_data.find(name);
+        return it != internals.shared_data.end() ? it->second : nullptr;
+    });
+}
+
+/// Set the shared data that can be later recovered by `get_shared_data()`.
+PYBIND11_NOINLINE void *set_shared_data(const std::string &name, void *data) {
+    return detail::with_internals([&](detail::internals &internals) {
+        internals.shared_data[name] = data;
+        return data;
+    });
+}
+
+/// Returns a typed reference to a shared data entry (by using `get_shared_data()`) if
+/// such entry exists. Otherwise, a new object of default-constructible type `T` is
+/// added to the shared data under the given name and a reference to it is returned.
+template <typename T>
+T &get_or_create_shared_data(const std::string &name) {
+    return *detail::with_internals([&](detail::internals &internals) {
+        auto it = internals.shared_data.find(name);
+        T *ptr = (T *) (it != internals.shared_data.end() ? it->second : nullptr);
+        if (!ptr) {
+            ptr = new T();
+            internals.shared_data[name] = ptr;
+        }
+        return ptr;
+    });
+}
+
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/detail/type_caster_base.h

@@ -0,0 +1,1195 @@
+/*
+    pybind11/detail/type_caster_base.h (originally first part of pybind11/cast.h)
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include <pybind11/pytypes.h>
+
+#include "common.h"
+#include "cpp_conduit.h"
+#include "descr.h"
+#include "internals.h"
+#include "typeid.h"
+#include "value_and_holder.h"
+
+#include <cstdint>
+#include <cstring>
+#include <iterator>
+#include <new>
+#include <stdexcept>
+#include <string>
+#include <type_traits>
+#include <typeindex>
+#include <typeinfo>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+/// A life support system for temporary objects created by `type_caster::load()`.
+/// Adding a patient will keep it alive up until the enclosing function returns.
+class loader_life_support {
+private:
+    loader_life_support *parent = nullptr;
+    std::unordered_set<PyObject *> keep_alive;
+
+    // Store stack pointer in thread-local storage.
+    static PYBIND11_TLS_KEY_REF get_stack_tls_key() {
+#if PYBIND11_INTERNALS_VERSION == 4
+        return get_local_internals().loader_life_support_tls_key;
+#else
+        return get_internals().loader_life_support_tls_key;
+#endif
+    }
+    static loader_life_support *get_stack_top() {
+        return static_cast<loader_life_support *>(PYBIND11_TLS_GET_VALUE(get_stack_tls_key()));
+    }
+    static void set_stack_top(loader_life_support *value) {
+        PYBIND11_TLS_REPLACE_VALUE(get_stack_tls_key(), value);
+    }
+
+public:
+    /// A new patient frame is created when a function is entered
+    loader_life_support() : parent{get_stack_top()} { set_stack_top(this); }
+
+    /// ... and destroyed after it returns
+    ~loader_life_support() {
+        if (get_stack_top() != this) {
+            pybind11_fail("loader_life_support: internal error");
+        }
+        set_stack_top(parent);
+        for (auto *item : keep_alive) {
+            Py_DECREF(item);
+        }
+    }
+
+    /// This can only be used inside a pybind11-bound function, either by `argument_loader`
+    /// at argument preparation time or by `py::cast()` at execution time.
+    PYBIND11_NOINLINE static void add_patient(handle h) {
+        loader_life_support *frame = get_stack_top();
+        if (!frame) {
+            // NOTE: It would be nice to include the stack frames here, as this indicates
+            // use of pybind11::cast<> outside the normal call framework, finding such
+            // a location is challenging. Developers could consider printing out
+            // stack frame addresses here using something like __builtin_frame_address(0)
+            throw cast_error("When called outside a bound function, py::cast() cannot "
+                             "do Python -> C++ conversions which require the creation "
+                             "of temporary values");
+        }
+
+        if (frame->keep_alive.insert(h.ptr()).second) {
+            Py_INCREF(h.ptr());
+        }
+    }
+};
+
+// Gets the cache entry for the given type, creating it if necessary.  The return value is the pair
+// returned by emplace, i.e. an iterator for the entry and a bool set to `true` if the entry was
+// just created.
+inline std::pair<decltype(internals::registered_types_py)::iterator, bool>
+all_type_info_get_cache(PyTypeObject *type);
+
+// Band-aid workaround to fix a subtle but serious bug in a minimalistic fashion. See PR #4762.
+inline void all_type_info_add_base_most_derived_first(std::vector<type_info *> &bases,
+                                                      type_info *addl_base) {
+    for (auto it = bases.begin(); it != bases.end(); it++) {
+        type_info *existing_base = *it;
+        if (PyType_IsSubtype(addl_base->type, existing_base->type) != 0) {
+            bases.insert(it, addl_base);
+            return;
+        }
+    }
+    bases.push_back(addl_base);
+}
+
+// Populates a just-created cache entry.
+PYBIND11_NOINLINE void all_type_info_populate(PyTypeObject *t, std::vector<type_info *> &bases) {
+    assert(bases.empty());
+    std::vector<PyTypeObject *> check;
+    for (handle parent : reinterpret_borrow<tuple>(t->tp_bases)) {
+        check.push_back((PyTypeObject *) parent.ptr());
+    }
+
+    auto const &type_dict = get_internals().registered_types_py;
+    for (size_t i = 0; i < check.size(); i++) {
+        auto *type = check[i];
+        // Ignore Python2 old-style class super type:
+        if (!PyType_Check((PyObject *) type)) {
+            continue;
+        }
+
+        // Check `type` in the current set of registered python types:
+        auto it = type_dict.find(type);
+        if (it != type_dict.end()) {
+            // We found a cache entry for it, so it's either pybind-registered or has pre-computed
+            // pybind bases, but we have to make sure we haven't already seen the type(s) before:
+            // we want to follow Python/virtual C++ rules that there should only be one instance of
+            // a common base.
+            for (auto *tinfo : it->second) {
+                // NB: Could use a second set here, rather than doing a linear search, but since
+                // having a large number of immediate pybind11-registered types seems fairly
+                // unlikely, that probably isn't worthwhile.
+                bool found = false;
+                for (auto *known : bases) {
+                    if (known == tinfo) {
+                        found = true;
+                        break;
+                    }
+                }
+                if (!found) {
+                    all_type_info_add_base_most_derived_first(bases, tinfo);
+                }
+            }
+        } else if (type->tp_bases) {
+            // It's some python type, so keep follow its bases classes to look for one or more
+            // registered types
+            if (i + 1 == check.size()) {
+                // When we're at the end, we can pop off the current element to avoid growing
+                // `check` when adding just one base (which is typical--i.e. when there is no
+                // multiple inheritance)
+                check.pop_back();
+                i--;
+            }
+            for (handle parent : reinterpret_borrow<tuple>(type->tp_bases)) {
+                check.push_back((PyTypeObject *) parent.ptr());
+            }
+        }
+    }
+}
+
+/**
+ * Extracts vector of type_info pointers of pybind-registered roots of the given Python type.  Will
+ * be just 1 pybind type for the Python type of a pybind-registered class, or for any Python-side
+ * derived class that uses single inheritance.  Will contain as many types as required for a Python
+ * class that uses multiple inheritance to inherit (directly or indirectly) from multiple
+ * pybind-registered classes.  Will be empty if neither the type nor any base classes are
+ * pybind-registered.
+ *
+ * The value is cached for the lifetime of the Python type.
+ */
+inline const std::vector<detail::type_info *> &all_type_info(PyTypeObject *type) {
+    auto ins = all_type_info_get_cache(type);
+    if (ins.second) {
+        // New cache entry: populate it
+        all_type_info_populate(type, ins.first->second);
+    }
+
+    return ins.first->second;
+}
+
+/**
+ * Gets a single pybind11 type info for a python type.  Returns nullptr if neither the type nor any
+ * ancestors are pybind11-registered.  Throws an exception if there are multiple bases--use
+ * `all_type_info` instead if you want to support multiple bases.
+ */
+PYBIND11_NOINLINE detail::type_info *get_type_info(PyTypeObject *type) {
+    const auto &bases = all_type_info(type);
+    if (bases.empty()) {
+        return nullptr;
+    }
+    if (bases.size() > 1) {
+        pybind11_fail(
+            "pybind11::detail::get_type_info: type has multiple pybind11-registered bases");
+    }
+    return bases.front();
+}
+
+inline detail::type_info *get_local_type_info(const std::type_index &tp) {
+    auto &locals = get_local_internals().registered_types_cpp;
+    auto it = locals.find(tp);
+    if (it != locals.end()) {
+        return it->second;
+    }
+    return nullptr;
+}
+
+inline detail::type_info *get_global_type_info(const std::type_index &tp) {
+    return with_internals([&](internals &internals) {
+        detail::type_info *type_info = nullptr;
+        auto &types = internals.registered_types_cpp;
+        auto it = types.find(tp);
+        if (it != types.end()) {
+            type_info = it->second;
+        }
+        return type_info;
+    });
+}
+
+/// Return the type info for a given C++ type; on lookup failure can either throw or return
+/// nullptr.
+PYBIND11_NOINLINE detail::type_info *get_type_info(const std::type_index &tp,
+                                                   bool throw_if_missing = false) {
+    if (auto *ltype = get_local_type_info(tp)) {
+        return ltype;
+    }
+    if (auto *gtype = get_global_type_info(tp)) {
+        return gtype;
+    }
+
+    if (throw_if_missing) {
+        std::string tname = tp.name();
+        detail::clean_type_id(tname);
+        pybind11_fail("pybind11::detail::get_type_info: unable to find type info for \""
+                      + std::move(tname) + '"');
+    }
+    return nullptr;
+}
+
+PYBIND11_NOINLINE handle get_type_handle(const std::type_info &tp, bool throw_if_missing) {
+    detail::type_info *type_info = get_type_info(tp, throw_if_missing);
+    return handle(type_info ? ((PyObject *) type_info->type) : nullptr);
+}
+
+// Searches the inheritance graph for a registered Python instance, using all_type_info().
+PYBIND11_NOINLINE handle find_registered_python_instance(void *src,
+                                                         const detail::type_info *tinfo) {
+    return with_instance_map(src, [&](instance_map &instances) {
+        auto it_instances = instances.equal_range(src);
+        for (auto it_i = it_instances.first; it_i != it_instances.second; ++it_i) {
+            for (auto *instance_type : detail::all_type_info(Py_TYPE(it_i->second))) {
+                if (instance_type && same_type(*instance_type->cpptype, *tinfo->cpptype)) {
+                    return handle((PyObject *) it_i->second).inc_ref();
+                }
+            }
+        }
+        return handle();
+    });
+}
+
+// Container for accessing and iterating over an instance's values/holders
+struct values_and_holders {
+private:
+    instance *inst;
+    using type_vec = std::vector<detail::type_info *>;
+    const type_vec &tinfo;
+
+public:
+    explicit values_and_holders(instance *inst)
+        : inst{inst}, tinfo(all_type_info(Py_TYPE(inst))) {}
+
+    explicit values_and_holders(PyObject *obj)
+        : inst{nullptr}, tinfo(all_type_info(Py_TYPE(obj))) {
+        if (!tinfo.empty()) {
+            inst = reinterpret_cast<instance *>(obj);
+        }
+    }
+
+    struct iterator {
+    private:
+        instance *inst = nullptr;
+        const type_vec *types = nullptr;
+        value_and_holder curr;
+        friend struct values_and_holders;
+        iterator(instance *inst, const type_vec *tinfo) : inst{inst}, types{tinfo} {
+            if (inst != nullptr) {
+                assert(!types->empty());
+                curr = value_and_holder(
+                    inst /* instance */,
+                    (*types)[0] /* type info */,
+                    0, /* vpos: (non-simple types only): the first vptr comes first */
+                    0 /* index */);
+            }
+        }
+        // Past-the-end iterator:
+        explicit iterator(size_t end) : curr(end) {}
+
+    public:
+        bool operator==(const iterator &other) const { return curr.index == other.curr.index; }
+        bool operator!=(const iterator &other) const { return curr.index != other.curr.index; }
+        iterator &operator++() {
+            if (!inst->simple_layout) {
+                curr.vh += 1 + (*types)[curr.index]->holder_size_in_ptrs;
+            }
+            ++curr.index;
+            curr.type = curr.index < types->size() ? (*types)[curr.index] : nullptr;
+            return *this;
+        }
+        value_and_holder &operator*() { return curr; }
+        value_and_holder *operator->() { return &curr; }
+    };
+
+    iterator begin() { return iterator(inst, &tinfo); }
+    iterator end() { return iterator(tinfo.size()); }
+
+    iterator find(const type_info *find_type) {
+        auto it = begin(), endit = end();
+        while (it != endit && it->type != find_type) {
+            ++it;
+        }
+        return it;
+    }
+
+    size_t size() { return tinfo.size(); }
+
+    // Band-aid workaround to fix a subtle but serious bug in a minimalistic fashion. See PR #4762.
+    bool is_redundant_value_and_holder(const value_and_holder &vh) {
+        for (size_t i = 0; i < vh.index; i++) {
+            if (PyType_IsSubtype(tinfo[i]->type, tinfo[vh.index]->type) != 0) {
+                return true;
+            }
+        }
+        return false;
+    }
+};
+
+/**
+ * Extracts C++ value and holder pointer references from an instance (which may contain multiple
+ * values/holders for python-side multiple inheritance) that match the given type.  Throws an error
+ * if the given type (or ValueType, if omitted) is not a pybind11 base of the given instance.  If
+ * `find_type` is omitted (or explicitly specified as nullptr) the first value/holder are returned,
+ * regardless of type (and the resulting .type will be nullptr).
+ *
+ * The returned object should be short-lived: in particular, it must not outlive the called-upon
+ * instance.
+ */
+PYBIND11_NOINLINE value_and_holder
+instance::get_value_and_holder(const type_info *find_type /*= nullptr default in common.h*/,
+                               bool throw_if_missing /*= true in common.h*/) {
+    // Optimize common case:
+    if (!find_type || Py_TYPE(this) == find_type->type) {
+        return value_and_holder(this, find_type, 0, 0);
+    }
+
+    detail::values_and_holders vhs(this);
+    auto it = vhs.find(find_type);
+    if (it != vhs.end()) {
+        return *it;
+    }
+
+    if (!throw_if_missing) {
+        return value_and_holder();
+    }
+
+#if defined(PYBIND11_DETAILED_ERROR_MESSAGES)
+    pybind11_fail("pybind11::detail::instance::get_value_and_holder: `"
+                  + get_fully_qualified_tp_name(find_type->type)
+                  + "' is not a pybind11 base of the given `"
+                  + get_fully_qualified_tp_name(Py_TYPE(this)) + "' instance");
+#else
+    pybind11_fail(
+        "pybind11::detail::instance::get_value_and_holder: "
+        "type is not a pybind11 base of the given instance "
+        "(#define PYBIND11_DETAILED_ERROR_MESSAGES or compile in debug mode for type details)");
+#endif
+}
+
+PYBIND11_NOINLINE void instance::allocate_layout() {
+    const auto &tinfo = all_type_info(Py_TYPE(this));
+
+    const size_t n_types = tinfo.size();
+
+    if (n_types == 0) {
+        pybind11_fail(
+            "instance allocation failed: new instance has no pybind11-registered base types");
+    }
+
+    simple_layout
+        = n_types == 1 && tinfo.front()->holder_size_in_ptrs <= instance_simple_holder_in_ptrs();
+
+    // Simple path: no python-side multiple inheritance, and a small-enough holder
+    if (simple_layout) {
+        simple_value_holder[0] = nullptr;
+        simple_holder_constructed = false;
+        simple_instance_registered = false;
+    } else { // multiple base types or a too-large holder
+        // Allocate space to hold: [v1*][h1][v2*][h2]...[bb...] where [vN*] is a value pointer,
+        // [hN] is the (uninitialized) holder instance for value N, and [bb...] is a set of bool
+        // values that tracks whether each associated holder has been initialized.  Each [block] is
+        // padded, if necessary, to an integer multiple of sizeof(void *).
+        size_t space = 0;
+        for (auto *t : tinfo) {
+            space += 1;                      // value pointer
+            space += t->holder_size_in_ptrs; // holder instance
+        }
+        size_t flags_at = space;
+        space += size_in_ptrs(n_types); // status bytes (holder_constructed and
+                                        // instance_registered)
+
+        // Allocate space for flags, values, and holders, and initialize it to 0 (flags and values,
+        // in particular, need to be 0).  Use Python's memory allocation
+        // functions: Python is using pymalloc, which is designed to be
+        // efficient for small allocations like the one we're doing here;
+        // for larger allocations they are just wrappers around malloc.
+        // TODO: is this still true for pure Python 3.6?
+        nonsimple.values_and_holders = (void **) PyMem_Calloc(space, sizeof(void *));
+        if (!nonsimple.values_and_holders) {
+            throw std::bad_alloc();
+        }
+        nonsimple.status
+            = reinterpret_cast<std::uint8_t *>(&nonsimple.values_and_holders[flags_at]);
+    }
+    owned = true;
+}
+
+// NOLINTNEXTLINE(readability-make-member-function-const)
+PYBIND11_NOINLINE void instance::deallocate_layout() {
+    if (!simple_layout) {
+        PyMem_Free(reinterpret_cast<void *>(nonsimple.values_and_holders));
+    }
+}
+
+PYBIND11_NOINLINE bool isinstance_generic(handle obj, const std::type_info &tp) {
+    handle type = detail::get_type_handle(tp, false);
+    if (!type) {
+        return false;
+    }
+    return isinstance(obj, type);
+}
+
+PYBIND11_NOINLINE handle get_object_handle(const void *ptr, const detail::type_info *type) {
+    return with_instance_map(ptr, [&](instance_map &instances) {
+        auto range = instances.equal_range(ptr);
+        for (auto it = range.first; it != range.second; ++it) {
+            for (const auto &vh : values_and_holders(it->second)) {
+                if (vh.type == type) {
+                    return handle((PyObject *) it->second);
+                }
+            }
+        }
+        return handle();
+    });
+}
+
+inline PyThreadState *get_thread_state_unchecked() {
+#if defined(PYPY_VERSION)
+    return PyThreadState_GET();
+#elif PY_VERSION_HEX < 0x030D0000
+    return _PyThreadState_UncheckedGet();
+#else
+    return PyThreadState_GetUnchecked();
+#endif
+}
+
+// Forward declarations
+void keep_alive_impl(handle nurse, handle patient);
+inline PyObject *make_new_instance(PyTypeObject *type);
+
+class type_caster_generic {
+public:
+    PYBIND11_NOINLINE explicit type_caster_generic(const std::type_info &type_info)
+        : typeinfo(get_type_info(type_info)), cpptype(&type_info) {}
+
+    explicit type_caster_generic(const type_info *typeinfo)
+        : typeinfo(typeinfo), cpptype(typeinfo ? typeinfo->cpptype : nullptr) {}
+
+    bool load(handle src, bool convert) { return load_impl<type_caster_generic>(src, convert); }
+
+    PYBIND11_NOINLINE static handle cast(const void *_src,
+                                         return_value_policy policy,
+                                         handle parent,
+                                         const detail::type_info *tinfo,
+                                         void *(*copy_constructor)(const void *),
+                                         void *(*move_constructor)(const void *),
+                                         const void *existing_holder = nullptr) {
+        if (!tinfo) { // no type info: error will be set already
+            return handle();
+        }
+
+        void *src = const_cast<void *>(_src);
+        if (src == nullptr) {
+            return none().release();
+        }
+
+        if (handle registered_inst = find_registered_python_instance(src, tinfo)) {
+            return registered_inst;
+        }
+
+        auto inst = reinterpret_steal<object>(make_new_instance(tinfo->type));
+        auto *wrapper = reinterpret_cast<instance *>(inst.ptr());
+        wrapper->owned = false;
+        void *&valueptr = values_and_holders(wrapper).begin()->value_ptr();
+
+        switch (policy) {
+            case return_value_policy::automatic:
+            case return_value_policy::take_ownership:
+                valueptr = src;
+                wrapper->owned = true;
+                break;
+
+            case return_value_policy::automatic_reference:
+            case return_value_policy::reference:
+                valueptr = src;
+                wrapper->owned = false;
+                break;
+
+            case return_value_policy::copy:
+                if (copy_constructor) {
+                    valueptr = copy_constructor(src);
+                } else {
+#if defined(PYBIND11_DETAILED_ERROR_MESSAGES)
+                    std::string type_name(tinfo->cpptype->name());
+                    detail::clean_type_id(type_name);
+                    throw cast_error("return_value_policy = copy, but type " + type_name
+                                     + " is non-copyable!");
+#else
+                    throw cast_error("return_value_policy = copy, but type is "
+                                     "non-copyable! (#define PYBIND11_DETAILED_ERROR_MESSAGES or "
+                                     "compile in debug mode for details)");
+#endif
+                }
+                wrapper->owned = true;
+                break;
+
+            case return_value_policy::move:
+                if (move_constructor) {
+                    valueptr = move_constructor(src);
+                } else if (copy_constructor) {
+                    valueptr = copy_constructor(src);
+                } else {
+#if defined(PYBIND11_DETAILED_ERROR_MESSAGES)
+                    std::string type_name(tinfo->cpptype->name());
+                    detail::clean_type_id(type_name);
+                    throw cast_error("return_value_policy = move, but type " + type_name
+                                     + " is neither movable nor copyable!");
+#else
+                    throw cast_error("return_value_policy = move, but type is neither "
+                                     "movable nor copyable! "
+                                     "(#define PYBIND11_DETAILED_ERROR_MESSAGES or compile in "
+                                     "debug mode for details)");
+#endif
+                }
+                wrapper->owned = true;
+                break;
+
+            case return_value_policy::reference_internal:
+                valueptr = src;
+                wrapper->owned = false;
+                keep_alive_impl(inst, parent);
+                break;
+
+            default:
+                throw cast_error("unhandled return_value_policy: should not happen!");
+        }
+
+        tinfo->init_instance(wrapper, existing_holder);
+
+        return inst.release();
+    }
+
+    // Base methods for generic caster; there are overridden in copyable_holder_caster
+    void load_value(value_and_holder &&v_h) {
+        auto *&vptr = v_h.value_ptr();
+        // Lazy allocation for unallocated values:
+        if (vptr == nullptr) {
+            const auto *type = v_h.type ? v_h.type : typeinfo;
+            if (type->operator_new) {
+                vptr = type->operator_new(type->type_size);
+            } else {
+#if defined(__cpp_aligned_new) && (!defined(_MSC_VER) || _MSC_VER >= 1912)
+                if (type->type_align > __STDCPP_DEFAULT_NEW_ALIGNMENT__) {
+                    vptr = ::operator new(type->type_size, std::align_val_t(type->type_align));
+                } else {
+                    vptr = ::operator new(type->type_size);
+                }
+#else
+                vptr = ::operator new(type->type_size);
+#endif
+            }
+        }
+        value = vptr;
+    }
+    bool try_implicit_casts(handle src, bool convert) {
+        for (const auto &cast : typeinfo->implicit_casts) {
+            type_caster_generic sub_caster(*cast.first);
+            if (sub_caster.load(src, convert)) {
+                value = cast.second(sub_caster.value);
+                return true;
+            }
+        }
+        return false;
+    }
+    bool try_direct_conversions(handle src) {
+        for (auto &converter : *typeinfo->direct_conversions) {
+            if (converter(src.ptr(), value)) {
+                return true;
+            }
+        }
+        return false;
+    }
+    bool try_cpp_conduit(handle src) {
+        value = try_raw_pointer_ephemeral_from_cpp_conduit(src, cpptype);
+        if (value != nullptr) {
+            return true;
+        }
+        return false;
+    }
+    void check_holder_compat() {}
+
+    PYBIND11_NOINLINE static void *local_load(PyObject *src, const type_info *ti) {
+        auto caster = type_caster_generic(ti);
+        if (caster.load(src, false)) {
+            return caster.value;
+        }
+        return nullptr;
+    }
+
+    /// Try to load with foreign typeinfo, if available. Used when there is no
+    /// native typeinfo, or when the native one wasn't able to produce a value.
+    PYBIND11_NOINLINE bool try_load_foreign_module_local(handle src) {
+        constexpr auto *local_key = PYBIND11_MODULE_LOCAL_ID;
+        const auto pytype = type::handle_of(src);
+        if (!hasattr(pytype, local_key)) {
+            return false;
+        }
+
+        type_info *foreign_typeinfo = reinterpret_borrow<capsule>(getattr(pytype, local_key));
+        // Only consider this foreign loader if actually foreign and is a loader of the correct cpp
+        // type
+        if (foreign_typeinfo->module_local_load == &local_load
+            || (cpptype && !same_type(*cpptype, *foreign_typeinfo->cpptype))) {
+            return false;
+        }
+
+        if (auto *result = foreign_typeinfo->module_local_load(src.ptr(), foreign_typeinfo)) {
+            value = result;
+            return true;
+        }
+        return false;
+    }
+
+    // Implementation of `load`; this takes the type of `this` so that it can dispatch the relevant
+    // bits of code between here and copyable_holder_caster where the two classes need different
+    // logic (without having to resort to virtual inheritance).
+    template <typename ThisT>
+    PYBIND11_NOINLINE bool load_impl(handle src, bool convert) {
+        if (!src) {
+            return false;
+        }
+        if (!typeinfo) {
+            return try_load_foreign_module_local(src);
+        }
+
+        auto &this_ = static_cast<ThisT &>(*this);
+        this_.check_holder_compat();
+
+        PyTypeObject *srctype = Py_TYPE(src.ptr());
+
+        // Case 1: If src is an exact type match for the target type then we can reinterpret_cast
+        // the instance's value pointer to the target type:
+        if (srctype == typeinfo->type) {
+            this_.load_value(reinterpret_cast<instance *>(src.ptr())->get_value_and_holder());
+            return true;
+        }
+        // Case 2: We have a derived class
+        if (PyType_IsSubtype(srctype, typeinfo->type)) {
+            const auto &bases = all_type_info(srctype);
+            bool no_cpp_mi = typeinfo->simple_type;
+
+            // Case 2a: the python type is a Python-inherited derived class that inherits from just
+            // one simple (no MI) pybind11 class, or is an exact match, so the C++ instance is of
+            // the right type and we can use reinterpret_cast.
+            // (This is essentially the same as case 2b, but because not using multiple inheritance
+            // is extremely common, we handle it specially to avoid the loop iterator and type
+            // pointer lookup overhead)
+            if (bases.size() == 1 && (no_cpp_mi || bases.front()->type == typeinfo->type)) {
+                this_.load_value(reinterpret_cast<instance *>(src.ptr())->get_value_and_holder());
+                return true;
+            }
+            // Case 2b: the python type inherits from multiple C++ bases.  Check the bases to see
+            // if we can find an exact match (or, for a simple C++ type, an inherited match); if
+            // so, we can safely reinterpret_cast to the relevant pointer.
+            if (bases.size() > 1) {
+                for (auto *base : bases) {
+                    if (no_cpp_mi ? PyType_IsSubtype(base->type, typeinfo->type)
+                                  : base->type == typeinfo->type) {
+                        this_.load_value(
+                            reinterpret_cast<instance *>(src.ptr())->get_value_and_holder(base));
+                        return true;
+                    }
+                }
+            }
+
+            // Case 2c: C++ multiple inheritance is involved and we couldn't find an exact type
+            // match in the registered bases, above, so try implicit casting (needed for proper C++
+            // casting when MI is involved).
+            if (this_.try_implicit_casts(src, convert)) {
+                return true;
+            }
+        }
+
+        // Perform an implicit conversion
+        if (convert) {
+            for (const auto &converter : typeinfo->implicit_conversions) {
+                auto temp = reinterpret_steal<object>(converter(src.ptr(), typeinfo->type));
+                if (load_impl<ThisT>(temp, false)) {
+                    loader_life_support::add_patient(temp);
+                    return true;
+                }
+            }
+            if (this_.try_direct_conversions(src)) {
+                return true;
+            }
+        }
+
+        // Failed to match local typeinfo. Try again with global.
+        if (typeinfo->module_local) {
+            if (auto *gtype = get_global_type_info(*typeinfo->cpptype)) {
+                typeinfo = gtype;
+                return load(src, false);
+            }
+        }
+
+        // Global typeinfo has precedence over foreign module_local
+        if (try_load_foreign_module_local(src)) {
+            return true;
+        }
+
+        // Custom converters didn't take None, now we convert None to nullptr.
+        if (src.is_none()) {
+            // Defer accepting None to other overloads (if we aren't in convert mode):
+            if (!convert) {
+                return false;
+            }
+            value = nullptr;
+            return true;
+        }
+
+        if (convert && cpptype && this_.try_cpp_conduit(src)) {
+            return true;
+        }
+
+        return false;
+    }
+
+    // Called to do type lookup and wrap the pointer and type in a pair when a dynamic_cast
+    // isn't needed or can't be used.  If the type is unknown, sets the error and returns a pair
+    // with .second = nullptr.  (p.first = nullptr is not an error: it becomes None).
+    PYBIND11_NOINLINE static std::pair<const void *, const type_info *>
+    src_and_type(const void *src,
+                 const std::type_info &cast_type,
+                 const std::type_info *rtti_type = nullptr) {
+        if (auto *tpi = get_type_info(cast_type)) {
+            return {src, const_cast<const type_info *>(tpi)};
+        }
+
+        // Not found, set error:
+        std::string tname = rtti_type ? rtti_type->name() : cast_type.name();
+        detail::clean_type_id(tname);
+        std::string msg = "Unregistered type : " + tname;
+        set_error(PyExc_TypeError, msg.c_str());
+        return {nullptr, nullptr};
+    }
+
+    const type_info *typeinfo = nullptr;
+    const std::type_info *cpptype = nullptr;
+    void *value = nullptr;
+};
+
+inline object cpp_conduit_method(handle self,
+                                 const bytes &pybind11_platform_abi_id,
+                                 const capsule &cpp_type_info_capsule,
+                                 const bytes &pointer_kind) {
+#ifdef PYBIND11_HAS_STRING_VIEW
+    using cpp_str = std::string_view;
+#else
+    using cpp_str = std::string;
+#endif
+    if (cpp_str(pybind11_platform_abi_id) != PYBIND11_PLATFORM_ABI_ID) {
+        return none();
+    }
+    if (std::strcmp(cpp_type_info_capsule.name(), typeid(std::type_info).name()) != 0) {
+        return none();
+    }
+    if (cpp_str(pointer_kind) != "raw_pointer_ephemeral") {
+        throw std::runtime_error("Invalid pointer_kind: \"" + std::string(pointer_kind) + "\"");
+    }
+    const auto *cpp_type_info = cpp_type_info_capsule.get_pointer<const std::type_info>();
+    type_caster_generic caster(*cpp_type_info);
+    if (!caster.load(self, false)) {
+        return none();
+    }
+    return capsule(caster.value, cpp_type_info->name());
+}
+
+/**
+ * Determine suitable casting operator for pointer-or-lvalue-casting type casters.  The type caster
+ * needs to provide `operator T*()` and `operator T&()` operators.
+ *
+ * If the type supports moving the value away via an `operator T&&() &&` method, it should use
+ * `movable_cast_op_type` instead.
+ */
+template <typename T>
+using cast_op_type = conditional_t<std::is_pointer<remove_reference_t<T>>::value,
+                                   typename std::add_pointer<intrinsic_t<T>>::type,
+                                   typename std::add_lvalue_reference<intrinsic_t<T>>::type>;
+
+/**
+ * Determine suitable casting operator for a type caster with a movable value.  Such a type caster
+ * needs to provide `operator T*()`, `operator T&()`, and `operator T&&() &&`.  The latter will be
+ * called in appropriate contexts where the value can be moved rather than copied.
+ *
+ * These operator are automatically provided when using the PYBIND11_TYPE_CASTER macro.
+ */
+template <typename T>
+using movable_cast_op_type
+    = conditional_t<std::is_pointer<typename std::remove_reference<T>::type>::value,
+                    typename std::add_pointer<intrinsic_t<T>>::type,
+                    conditional_t<std::is_rvalue_reference<T>::value,
+                                  typename std::add_rvalue_reference<intrinsic_t<T>>::type,
+                                  typename std::add_lvalue_reference<intrinsic_t<T>>::type>>;
+
+// Does the container have a mapped type and is it recursive?
+// Implemented by specializations below.
+template <typename Container, typename SFINAE = void>
+struct container_mapped_type_traits {
+    static constexpr bool has_mapped_type = false;
+    static constexpr bool has_recursive_mapped_type = false;
+};
+
+template <typename Container>
+struct container_mapped_type_traits<
+    Container,
+    typename std::enable_if<
+        std::is_same<typename Container::mapped_type, Container>::value>::type> {
+    static constexpr bool has_mapped_type = true;
+    static constexpr bool has_recursive_mapped_type = true;
+};
+
+template <typename Container>
+struct container_mapped_type_traits<
+    Container,
+    typename std::enable_if<
+        negation<std::is_same<typename Container::mapped_type, Container>>::value>::type> {
+    static constexpr bool has_mapped_type = true;
+    static constexpr bool has_recursive_mapped_type = false;
+};
+
+// Does the container have a value type and is it recursive?
+// Implemented by specializations below.
+template <typename Container, typename SFINAE = void>
+struct container_value_type_traits : std::false_type {
+    static constexpr bool has_value_type = false;
+    static constexpr bool has_recursive_value_type = false;
+};
+
+template <typename Container>
+struct container_value_type_traits<
+    Container,
+    typename std::enable_if<
+        std::is_same<typename Container::value_type, Container>::value>::type> {
+    static constexpr bool has_value_type = true;
+    static constexpr bool has_recursive_value_type = true;
+};
+
+template <typename Container>
+struct container_value_type_traits<
+    Container,
+    typename std::enable_if<
+        negation<std::is_same<typename Container::value_type, Container>>::value>::type> {
+    static constexpr bool has_value_type = true;
+    static constexpr bool has_recursive_value_type = false;
+};
+
+/*
+ * Tag to be used for representing the bottom of recursively defined types.
+ * Define this tag so we don't have to use void.
+ */
+struct recursive_bottom {};
+
+/*
+ * Implementation detail of `recursive_container_traits` below.
+ * `T` is the `value_type` of the container, which might need to be modified to
+ * avoid recursive types and const types.
+ */
+template <typename T, bool is_this_a_map>
+struct impl_type_to_check_recursively {
+    /*
+     * If the container is recursive, then no further recursion should be done.
+     */
+    using if_recursive = recursive_bottom;
+    /*
+     * Otherwise yield `T` unchanged.
+     */
+    using if_not_recursive = T;
+};
+
+/*
+ * For pairs - only as value type of a map -, the first type should remove the `const`.
+ * Also, if the map is recursive, then the recursive checking should consider
+ * the first type only.
+ */
+template <typename A, typename B>
+struct impl_type_to_check_recursively<std::pair<A, B>, /* is_this_a_map = */ true> {
+    using if_recursive = typename std::remove_const<A>::type;
+    using if_not_recursive = std::pair<typename std::remove_const<A>::type, B>;
+};
+
+/*
+ * Implementation of `recursive_container_traits` below.
+ */
+template <typename Container, typename SFINAE = void>
+struct impl_recursive_container_traits {
+    using type_to_check_recursively = recursive_bottom;
+};
+
+template <typename Container>
+struct impl_recursive_container_traits<
+    Container,
+    typename std::enable_if<container_value_type_traits<Container>::has_value_type>::type> {
+    static constexpr bool is_recursive
+        = container_mapped_type_traits<Container>::has_recursive_mapped_type
+          || container_value_type_traits<Container>::has_recursive_value_type;
+    /*
+     * This member dictates which type Pybind11 should check recursively in traits
+     * such as `is_move_constructible`, `is_copy_constructible`, `is_move_assignable`, ...
+     * Direct access to `value_type` should be avoided:
+     * 1. `value_type` might recursively contain the type again
+     * 2. `value_type` of STL map types is `std::pair<A const, B>`, the `const`
+     *    should be removed.
+     *
+     */
+    using type_to_check_recursively = typename std::conditional<
+        is_recursive,
+        typename impl_type_to_check_recursively<
+            typename Container::value_type,
+            container_mapped_type_traits<Container>::has_mapped_type>::if_recursive,
+        typename impl_type_to_check_recursively<
+            typename Container::value_type,
+            container_mapped_type_traits<Container>::has_mapped_type>::if_not_recursive>::type;
+};
+
+/*
+ * This trait defines the `type_to_check_recursively` which is needed to properly
+ * handle recursively defined traits such as `is_move_constructible` without going
+ * into an infinite recursion.
+ * Should be used instead of directly accessing the `value_type`.
+ * It cancels the recursion by returning the `recursive_bottom` tag.
+ *
+ * The default definition of `type_to_check_recursively` is as follows:
+ *
+ * 1. By default, it is `recursive_bottom`, so that the recursion is canceled.
+ * 2. If the type is non-recursive and defines a `value_type`, then the `value_type` is used.
+ *    If the `value_type` is a pair and a `mapped_type` is defined,
+ *    then the `const` is removed from the first type.
+ * 3. If the type is recursive and `value_type` is not a pair, then `recursive_bottom` is returned.
+ * 4. If the type is recursive and `value_type` is a pair and a `mapped_type` is defined,
+ *    then `const` is removed from the first type and the first type is returned.
+ *
+ * This behavior can be extended by the user as seen in test_stl_binders.cpp.
+ *
+ * This struct is exactly the same as impl_recursive_container_traits.
+ * The duplication achieves that user-defined specializations don't compete
+ * with internal specializations, but take precedence.
+ */
+template <typename Container, typename SFINAE = void>
+struct recursive_container_traits : impl_recursive_container_traits<Container> {};
+
+template <typename T>
+struct is_move_constructible
+    : all_of<std::is_move_constructible<T>,
+             is_move_constructible<
+                 typename recursive_container_traits<T>::type_to_check_recursively>> {};
+
+template <>
+struct is_move_constructible<recursive_bottom> : std::true_type {};
+
+// Likewise for std::pair
+// (after C++17 it is mandatory that the move constructor not exist when the two types aren't
+// themselves move constructible, but this can not be relied upon when T1 or T2 are themselves
+// containers).
+template <typename T1, typename T2>
+struct is_move_constructible<std::pair<T1, T2>>
+    : all_of<is_move_constructible<T1>, is_move_constructible<T2>> {};
+
+// std::is_copy_constructible isn't quite enough: it lets std::vector<T> (and similar) through when
+// T is non-copyable, but code containing such a copy constructor fails to actually compile.
+template <typename T>
+struct is_copy_constructible
+    : all_of<std::is_copy_constructible<T>,
+             is_copy_constructible<
+                 typename recursive_container_traits<T>::type_to_check_recursively>> {};
+
+template <>
+struct is_copy_constructible<recursive_bottom> : std::true_type {};
+
+// Likewise for std::pair
+// (after C++17 it is mandatory that the copy constructor not exist when the two types aren't
+// themselves copy constructible, but this can not be relied upon when T1 or T2 are themselves
+// containers).
+template <typename T1, typename T2>
+struct is_copy_constructible<std::pair<T1, T2>>
+    : all_of<is_copy_constructible<T1>, is_copy_constructible<T2>> {};
+
+// The same problems arise with std::is_copy_assignable, so we use the same workaround.
+template <typename T>
+struct is_copy_assignable
+    : all_of<
+          std::is_copy_assignable<T>,
+          is_copy_assignable<typename recursive_container_traits<T>::type_to_check_recursively>> {
+};
+
+template <>
+struct is_copy_assignable<recursive_bottom> : std::true_type {};
+
+template <typename T1, typename T2>
+struct is_copy_assignable<std::pair<T1, T2>>
+    : all_of<is_copy_assignable<T1>, is_copy_assignable<T2>> {};
+
+PYBIND11_NAMESPACE_END(detail)
+
+// polymorphic_type_hook<itype>::get(src, tinfo) determines whether the object pointed
+// to by `src` actually is an instance of some class derived from `itype`.
+// If so, it sets `tinfo` to point to the std::type_info representing that derived
+// type, and returns a pointer to the start of the most-derived object of that type
+// (in which `src` is a subobject; this will be the same address as `src` in most
+// single inheritance cases). If not, or if `src` is nullptr, it simply returns `src`
+// and leaves `tinfo` at its default value of nullptr.
+//
+// The default polymorphic_type_hook just returns src. A specialization for polymorphic
+// types determines the runtime type of the passed object and adjusts the this-pointer
+// appropriately via dynamic_cast<void*>. This is what enables a C++ Animal* to appear
+// to Python as a Dog (if Dog inherits from Animal, Animal is polymorphic, Dog is
+// registered with pybind11, and this Animal is in fact a Dog).
+//
+// You may specialize polymorphic_type_hook yourself for types that want to appear
+// polymorphic to Python but do not use C++ RTTI. (This is a not uncommon pattern
+// in performance-sensitive applications, used most notably in LLVM.)
+//
+// polymorphic_type_hook_base allows users to specialize polymorphic_type_hook with
+// std::enable_if. User provided specializations will always have higher priority than
+// the default implementation and specialization provided in polymorphic_type_hook_base.
+template <typename itype, typename SFINAE = void>
+struct polymorphic_type_hook_base {
+    static const void *get(const itype *src, const std::type_info *&) { return src; }
+};
+template <typename itype>
+struct polymorphic_type_hook_base<itype, detail::enable_if_t<std::is_polymorphic<itype>::value>> {
+    static const void *get(const itype *src, const std::type_info *&type) {
+        type = src ? &typeid(*src) : nullptr;
+        return dynamic_cast<const void *>(src);
+    }
+};
+template <typename itype, typename SFINAE = void>
+struct polymorphic_type_hook : public polymorphic_type_hook_base<itype> {};
+
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+/// Generic type caster for objects stored on the heap
+template <typename type>
+class type_caster_base : public type_caster_generic {
+    using itype = intrinsic_t<type>;
+
+public:
+    static constexpr auto name = const_name<type>();
+
+    type_caster_base() : type_caster_base(typeid(type)) {}
+    explicit type_caster_base(const std::type_info &info) : type_caster_generic(info) {}
+
+    static handle cast(const itype &src, return_value_policy policy, handle parent) {
+        if (policy == return_value_policy::automatic
+            || policy == return_value_policy::automatic_reference) {
+            policy = return_value_policy::copy;
+        }
+        return cast(std::addressof(src), policy, parent);
+    }
+
+    static handle cast(itype &&src, return_value_policy, handle parent) {
+        return cast(std::addressof(src), return_value_policy::move, parent);
+    }
+
+    // Returns a (pointer, type_info) pair taking care of necessary type lookup for a
+    // polymorphic type (using RTTI by default, but can be overridden by specializing
+    // polymorphic_type_hook). If the instance isn't derived, returns the base version.
+    static std::pair<const void *, const type_info *> src_and_type(const itype *src) {
+        const auto &cast_type = typeid(itype);
+        const std::type_info *instance_type = nullptr;
+        const void *vsrc = polymorphic_type_hook<itype>::get(src, instance_type);
+        if (instance_type && !same_type(cast_type, *instance_type)) {
+            // This is a base pointer to a derived type. If the derived type is registered
+            // with pybind11, we want to make the full derived object available.
+            // In the typical case where itype is polymorphic, we get the correct
+            // derived pointer (which may be != base pointer) by a dynamic_cast to
+            // most derived type. If itype is not polymorphic, we won't get here
+            // except via a user-provided specialization of polymorphic_type_hook,
+            // and the user has promised that no this-pointer adjustment is
+            // required in that case, so it's OK to use static_cast.
+            if (const auto *tpi = get_type_info(*instance_type)) {
+                return {vsrc, tpi};
+            }
+        }
+        // Otherwise we have either a nullptr, an `itype` pointer, or an unknown derived pointer,
+        // so don't do a cast
+        return type_caster_generic::src_and_type(src, cast_type, instance_type);
+    }
+
+    static handle cast(const itype *src, return_value_policy policy, handle parent) {
+        auto st = src_and_type(src);
+        return type_caster_generic::cast(st.first,
+                                         policy,
+                                         parent,
+                                         st.second,
+                                         make_copy_constructor(src),
+                                         make_move_constructor(src));
+    }
+
+    static handle cast_holder(const itype *src, const void *holder) {
+        auto st = src_and_type(src);
+        return type_caster_generic::cast(st.first,
+                                         return_value_policy::take_ownership,
+                                         {},
+                                         st.second,
+                                         nullptr,
+                                         nullptr,
+                                         holder);
+    }
+
+    template <typename T>
+    using cast_op_type = detail::cast_op_type<T>;
+
+    // NOLINTNEXTLINE(google-explicit-constructor)
+    operator itype *() { return (type *) value; }
+    // NOLINTNEXTLINE(google-explicit-constructor)
+    operator itype &() {
+        if (!value) {
+            throw reference_cast_error();
+        }
+        return *((itype *) value);
+    }
+
+protected:
+    using Constructor = void *(*) (const void *);
+
+    /* Only enabled when the types are {copy,move}-constructible *and* when the type
+       does not have a private operator new implementation. A comma operator is used in the
+       decltype argument to apply SFINAE to the public copy/move constructors.*/
+    template <typename T, typename = enable_if_t<is_copy_constructible<T>::value>>
+    static auto make_copy_constructor(const T *) -> decltype(new T(std::declval<const T>()),
+                                                             Constructor{}) {
+        return [](const void *arg) -> void * { return new T(*reinterpret_cast<const T *>(arg)); };
+    }
+
+    template <typename T, typename = enable_if_t<is_move_constructible<T>::value>>
+    static auto make_move_constructor(const T *) -> decltype(new T(std::declval<T &&>()),
+                                                             Constructor{}) {
+        return [](const void *arg) -> void * {
+            return new T(std::move(*const_cast<T *>(reinterpret_cast<const T *>(arg))));
+        };
+    }
+
+    static Constructor make_copy_constructor(...) { return nullptr; }
+    static Constructor make_move_constructor(...) { return nullptr; }
+};
+
+inline std::string quote_cpp_type_name(const std::string &cpp_type_name) {
+    return cpp_type_name; // No-op for now. See PR #4888
+}
+
+PYBIND11_NOINLINE std::string type_info_description(const std::type_info &ti) {
+    if (auto *type_data = get_type_info(ti)) {
+        handle th((PyObject *) type_data->type);
+        return th.attr("__module__").cast<std::string>() + '.'
+               + th.attr("__qualname__").cast<std::string>();
+    }
+    return quote_cpp_type_name(clean_type_id(ti.name()));
+}
+
+PYBIND11_NAMESPACE_END(detail)
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/detail/typeid.h

@@ -0,0 +1,65 @@
+/*
+    pybind11/detail/typeid.h: Compiler-independent access to type identifiers
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include <cstdio>
+#include <cstdlib>
+
+#if defined(__GNUG__)
+#    include <cxxabi.h>
+#endif
+
+#include "common.h"
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+/// Erase all occurrences of a substring
+inline void erase_all(std::string &string, const std::string &search) {
+    for (size_t pos = 0;;) {
+        pos = string.find(search, pos);
+        if (pos == std::string::npos) {
+            break;
+        }
+        string.erase(pos, search.length());
+    }
+}
+
+PYBIND11_NOINLINE void clean_type_id(std::string &name) {
+#if defined(__GNUG__)
+    int status = 0;
+    std::unique_ptr<char, void (*)(void *)> res{
+        abi::__cxa_demangle(name.c_str(), nullptr, nullptr, &status), std::free};
+    if (status == 0) {
+        name = res.get();
+    }
+#else
+    detail::erase_all(name, "class ");
+    detail::erase_all(name, "struct ");
+    detail::erase_all(name, "enum ");
+#endif
+    detail::erase_all(name, "pybind11::");
+}
+
+inline std::string clean_type_id(const char *typeid_name) {
+    std::string name(typeid_name);
+    detail::clean_type_id(name);
+    return name;
+}
+
+PYBIND11_NAMESPACE_END(detail)
+
+/// Return a string representation of a C++ type
+template <typename T>
+static std::string type_id() {
+    return detail::clean_type_id(typeid(T).name());
+}
+
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/detail/value_and_holder.h

@@ -0,0 +1,77 @@
+// Copyright (c) 2016-2024 The Pybind Development Team.
+// All rights reserved. Use of this source code is governed by a
+// BSD-style license that can be found in the LICENSE file.
+
+#pragma once
+
+#include "common.h"
+
+#include <cstddef>
+#include <typeinfo>
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+struct value_and_holder {
+    instance *inst = nullptr;
+    size_t index = 0u;
+    const detail::type_info *type = nullptr;
+    void **vh = nullptr;
+
+    // Main constructor for a found value/holder:
+    value_and_holder(instance *i, const detail::type_info *type, size_t vpos, size_t index)
+        : inst{i}, index{index}, type{type},
+          vh{inst->simple_layout ? inst->simple_value_holder
+                                 : &inst->nonsimple.values_and_holders[vpos]} {}
+
+    // Default constructor (used to signal a value-and-holder not found by get_value_and_holder())
+    value_and_holder() = default;
+
+    // Used for past-the-end iterator
+    explicit value_and_holder(size_t index) : index{index} {}
+
+    template <typename V = void>
+    V *&value_ptr() const {
+        return reinterpret_cast<V *&>(vh[0]);
+    }
+    // True if this `value_and_holder` has a non-null value pointer
+    explicit operator bool() const { return value_ptr() != nullptr; }
+
+    template <typename H>
+    H &holder() const {
+        return reinterpret_cast<H &>(vh[1]);
+    }
+    bool holder_constructed() const {
+        return inst->simple_layout
+                   ? inst->simple_holder_constructed
+                   : (inst->nonsimple.status[index] & instance::status_holder_constructed) != 0u;
+    }
+    // NOLINTNEXTLINE(readability-make-member-function-const)
+    void set_holder_constructed(bool v = true) {
+        if (inst->simple_layout) {
+            inst->simple_holder_constructed = v;
+        } else if (v) {
+            inst->nonsimple.status[index] |= instance::status_holder_constructed;
+        } else {
+            inst->nonsimple.status[index] &= (std::uint8_t) ~instance::status_holder_constructed;
+        }
+    }
+    bool instance_registered() const {
+        return inst->simple_layout
+                   ? inst->simple_instance_registered
+                   : ((inst->nonsimple.status[index] & instance::status_instance_registered) != 0);
+    }
+    // NOLINTNEXTLINE(readability-make-member-function-const)
+    void set_instance_registered(bool v = true) {
+        if (inst->simple_layout) {
+            inst->simple_instance_registered = v;
+        } else if (v) {
+            inst->nonsimple.status[index] |= instance::status_instance_registered;
+        } else {
+            inst->nonsimple.status[index] &= (std::uint8_t) ~instance::status_instance_registered;
+        }
+    }
+};
+
+PYBIND11_NAMESPACE_END(detail)
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/eigen.h

@@ -0,0 +1,12 @@
+/*
+    pybind11/eigen.h: Transparent conversion for dense and sparse Eigen matrices
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "eigen/matrix.h"

phivenv/Lib/site-packages/torch/include/pybind11/eigen/common.h

@@ -0,0 +1,9 @@
+// Copyright (c) 2023 The pybind Community.
+
+#pragma once
+
+// Common message for `static_assert()`s, which are useful to easily
+// preempt much less obvious errors.
+#define PYBIND11_EIGEN_MESSAGE_POINTER_TYPES_ARE_NOT_SUPPORTED                                    \
+    "Pointer types (in particular `PyObject *`) are not supported as scalar types for Eigen "     \
+    "types."

phivenv/Lib/site-packages/torch/include/pybind11/eigen/matrix.h

@@ -0,0 +1,715 @@
+/*
+    pybind11/eigen/matrix.h: Transparent conversion for dense and sparse Eigen matrices
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include <pybind11/numpy.h>
+
+#include "common.h"
+
+/* HINT: To suppress warnings originating from the Eigen headers, use -isystem.
+   See also:
+       https://stackoverflow.com/questions/2579576/i-dir-vs-isystem-dir
+       https://stackoverflow.com/questions/1741816/isystem-for-ms-visual-studio-c-compiler
+*/
+PYBIND11_WARNING_PUSH
+PYBIND11_WARNING_DISABLE_MSVC(5054) // https://github.com/pybind/pybind11/pull/3741
+//       C5054: operator '&': deprecated between enumerations of different types
+#if defined(__MINGW32__)
+PYBIND11_WARNING_DISABLE_GCC("-Wmaybe-uninitialized")
+#endif
+
+#include <Eigen/Core>
+#include <Eigen/SparseCore>
+
+PYBIND11_WARNING_POP
+
+// Eigen prior to 3.2.7 doesn't have proper move constructors--but worse, some classes get implicit
+// move constructors that break things.  We could detect this an explicitly copy, but an extra copy
+// of matrices seems highly undesirable.
+static_assert(EIGEN_VERSION_AT_LEAST(3, 2, 7),
+              "Eigen matrix support in pybind11 requires Eigen >= 3.2.7");
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+PYBIND11_WARNING_DISABLE_MSVC(4127)
+
+// Provide a convenience alias for easier pass-by-ref usage with fully dynamic strides:
+using EigenDStride = Eigen::Stride<Eigen::Dynamic, Eigen::Dynamic>;
+template <typename MatrixType>
+using EigenDRef = Eigen::Ref<MatrixType, 0, EigenDStride>;
+template <typename MatrixType>
+using EigenDMap = Eigen::Map<MatrixType, 0, EigenDStride>;
+
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+#if EIGEN_VERSION_AT_LEAST(3, 3, 0)
+using EigenIndex = Eigen::Index;
+template <typename Scalar, int Flags, typename StorageIndex>
+using EigenMapSparseMatrix = Eigen::Map<Eigen::SparseMatrix<Scalar, Flags, StorageIndex>>;
+#else
+using EigenIndex = EIGEN_DEFAULT_DENSE_INDEX_TYPE;
+template <typename Scalar, int Flags, typename StorageIndex>
+using EigenMapSparseMatrix = Eigen::MappedSparseMatrix<Scalar, Flags, StorageIndex>;
+#endif
+
+// Matches Eigen::Map, Eigen::Ref, blocks, etc:
+template <typename T>
+using is_eigen_dense_map = all_of<is_template_base_of<Eigen::DenseBase, T>,
+                                  std::is_base_of<Eigen::MapBase<T, Eigen::ReadOnlyAccessors>, T>>;
+template <typename T>
+using is_eigen_mutable_map = std::is_base_of<Eigen::MapBase<T, Eigen::WriteAccessors>, T>;
+template <typename T>
+using is_eigen_dense_plain
+    = all_of<negation<is_eigen_dense_map<T>>, is_template_base_of<Eigen::PlainObjectBase, T>>;
+template <typename T>
+using is_eigen_sparse = is_template_base_of<Eigen::SparseMatrixBase, T>;
+// Test for objects inheriting from EigenBase<Derived> that aren't captured by the above.  This
+// basically covers anything that can be assigned to a dense matrix but that don't have a typical
+// matrix data layout that can be copied from their .data().  For example, DiagonalMatrix and
+// SelfAdjointView fall into this category.
+template <typename T>
+using is_eigen_other
+    = all_of<is_template_base_of<Eigen::EigenBase, T>,
+             negation<any_of<is_eigen_dense_map<T>, is_eigen_dense_plain<T>, is_eigen_sparse<T>>>>;
+
+// Captures numpy/eigen conformability status (returned by EigenProps::conformable()):
+template <bool EigenRowMajor>
+struct EigenConformable {
+    bool conformable = false;
+    EigenIndex rows = 0, cols = 0;
+    EigenDStride stride{0, 0};    // Only valid if negativestrides is false!
+    bool negativestrides = false; // If true, do not use stride!
+
+    // NOLINTNEXTLINE(google-explicit-constructor)
+    EigenConformable(bool fits = false) : conformable{fits} {}
+    // Matrix type:
+    EigenConformable(EigenIndex r, EigenIndex c, EigenIndex rstride, EigenIndex cstride)
+        : conformable{true}, rows{r}, cols{c},
+          // TODO: when Eigen bug #747 is fixed, remove the tests for non-negativity.
+          // http://eigen.tuxfamily.org/bz/show_bug.cgi?id=747
+          stride{EigenRowMajor ? (rstride > 0 ? rstride : 0)
+                               : (cstride > 0 ? cstride : 0) /* outer stride */,
+                 EigenRowMajor ? (cstride > 0 ? cstride : 0)
+                               : (rstride > 0 ? rstride : 0) /* inner stride */},
+          negativestrides{rstride < 0 || cstride < 0} {}
+    // Vector type:
+    EigenConformable(EigenIndex r, EigenIndex c, EigenIndex stride)
+        : EigenConformable(r, c, r == 1 ? c * stride : stride, c == 1 ? r : r * stride) {}
+
+    template <typename props>
+    bool stride_compatible() const {
+        // To have compatible strides, we need (on both dimensions) one of fully dynamic strides,
+        // matching strides, or a dimension size of 1 (in which case the stride value is
+        // irrelevant). Alternatively, if any dimension size is 0, the strides are not relevant
+        // (and numpy ≥ 1.23 sets the strides to 0 in that case, so we need to check explicitly).
+        if (negativestrides) {
+            return false;
+        }
+        if (rows == 0 || cols == 0) {
+            return true;
+        }
+        return (props::inner_stride == Eigen::Dynamic || props::inner_stride == stride.inner()
+                || (EigenRowMajor ? cols : rows) == 1)
+               && (props::outer_stride == Eigen::Dynamic || props::outer_stride == stride.outer()
+                   || (EigenRowMajor ? rows : cols) == 1);
+    }
+    // NOLINTNEXTLINE(google-explicit-constructor)
+    operator bool() const { return conformable; }
+};
+
+template <typename Type>
+struct eigen_extract_stride {
+    using type = Type;
+};
+template <typename PlainObjectType, int MapOptions, typename StrideType>
+struct eigen_extract_stride<Eigen::Map<PlainObjectType, MapOptions, StrideType>> {
+    using type = StrideType;
+};
+template <typename PlainObjectType, int Options, typename StrideType>
+struct eigen_extract_stride<Eigen::Ref<PlainObjectType, Options, StrideType>> {
+    using type = StrideType;
+};
+
+// Helper struct for extracting information from an Eigen type
+template <typename Type_>
+struct EigenProps {
+    using Type = Type_;
+    using Scalar = typename Type::Scalar;
+    using StrideType = typename eigen_extract_stride<Type>::type;
+    static constexpr EigenIndex rows = Type::RowsAtCompileTime, cols = Type::ColsAtCompileTime,
+                                size = Type::SizeAtCompileTime;
+    static constexpr bool row_major = Type::IsRowMajor,
+                          vector
+                          = Type::IsVectorAtCompileTime, // At least one dimension has fixed size 1
+        fixed_rows = rows != Eigen::Dynamic, fixed_cols = cols != Eigen::Dynamic,
+                          fixed = size != Eigen::Dynamic, // Fully-fixed size
+        dynamic = !fixed_rows && !fixed_cols;             // Fully-dynamic size
+
+    template <EigenIndex i, EigenIndex ifzero>
+    using if_zero = std::integral_constant<EigenIndex, i == 0 ? ifzero : i>;
+    static constexpr EigenIndex inner_stride
+        = if_zero<StrideType::InnerStrideAtCompileTime, 1>::value,
+        outer_stride = if_zero < StrideType::OuterStrideAtCompileTime,
+        vector      ? size
+        : row_major ? cols
+                    : rows > ::value;
+    static constexpr bool dynamic_stride
+        = inner_stride == Eigen::Dynamic && outer_stride == Eigen::Dynamic;
+    static constexpr bool requires_row_major
+        = !dynamic_stride && !vector && (row_major ? inner_stride : outer_stride) == 1;
+    static constexpr bool requires_col_major
+        = !dynamic_stride && !vector && (row_major ? outer_stride : inner_stride) == 1;
+
+    // Takes an input array and determines whether we can make it fit into the Eigen type.  If
+    // the array is a vector, we attempt to fit it into either an Eigen 1xN or Nx1 vector
+    // (preferring the latter if it will fit in either, i.e. for a fully dynamic matrix type).
+    static EigenConformable<row_major> conformable(const array &a) {
+        const auto dims = a.ndim();
+        if (dims < 1 || dims > 2) {
+            return false;
+        }
+
+        if (dims == 2) { // Matrix type: require exact match (or dynamic)
+
+            EigenIndex np_rows = a.shape(0), np_cols = a.shape(1),
+                       np_rstride = a.strides(0) / static_cast<ssize_t>(sizeof(Scalar)),
+                       np_cstride = a.strides(1) / static_cast<ssize_t>(sizeof(Scalar));
+            if ((fixed_rows && np_rows != rows) || (fixed_cols && np_cols != cols)) {
+                return false;
+            }
+
+            return {np_rows, np_cols, np_rstride, np_cstride};
+        }
+
+        // Otherwise we're storing an n-vector.  Only one of the strides will be used, but
+        // whichever is used, we want the (single) numpy stride value.
+        const EigenIndex n = a.shape(0),
+                         stride = a.strides(0) / static_cast<ssize_t>(sizeof(Scalar));
+
+        if (vector) { // Eigen type is a compile-time vector
+            if (fixed && size != n) {
+                return false; // Vector size mismatch
+            }
+            return {rows == 1 ? 1 : n, cols == 1 ? 1 : n, stride};
+        }
+        if (fixed) {
+            // The type has a fixed size, but is not a vector: abort
+            return false;
+        }
+        if (fixed_cols) {
+            // Since this isn't a vector, cols must be != 1.  We allow this only if it exactly
+            // equals the number of elements (rows is Dynamic, and so 1 row is allowed).
+            if (cols != n) {
+                return false;
+            }
+            return {1, n, stride};
+        } // Otherwise it's either fully dynamic, or column dynamic; both become a column vector
+        if (fixed_rows && rows != n) {
+            return false;
+        }
+        return {n, 1, stride};
+    }
+
+    static constexpr bool show_writeable
+        = is_eigen_dense_map<Type>::value && is_eigen_mutable_map<Type>::value;
+    static constexpr bool show_order = is_eigen_dense_map<Type>::value;
+    static constexpr bool show_c_contiguous = show_order && requires_row_major;
+    static constexpr bool show_f_contiguous
+        = !show_c_contiguous && show_order && requires_col_major;
+
+    static constexpr auto descriptor
+        = const_name("numpy.ndarray[") + npy_format_descriptor<Scalar>::name + const_name("[")
+          + const_name<fixed_rows>(const_name<(size_t) rows>(), const_name("m")) + const_name(", ")
+          + const_name<fixed_cols>(const_name<(size_t) cols>(), const_name("n")) + const_name("]")
+          +
+          // For a reference type (e.g. Ref<MatrixXd>) we have other constraints that might need to
+          // be satisfied: writeable=True (for a mutable reference), and, depending on the map's
+          // stride options, possibly f_contiguous or c_contiguous.  We include them in the
+          // descriptor output to provide some hint as to why a TypeError is occurring (otherwise
+          // it can be confusing to see that a function accepts a 'numpy.ndarray[float64[3,2]]' and
+          // an error message that you *gave* a numpy.ndarray of the right type and dimensions.
+          const_name<show_writeable>(", flags.writeable", "")
+          + const_name<show_c_contiguous>(", flags.c_contiguous", "")
+          + const_name<show_f_contiguous>(", flags.f_contiguous", "") + const_name("]");
+};
+
+// Casts an Eigen type to numpy array.  If given a base, the numpy array references the src data,
+// otherwise it'll make a copy.  writeable lets you turn off the writeable flag for the array.
+template <typename props>
+handle
+eigen_array_cast(typename props::Type const &src, handle base = handle(), bool writeable = true) {
+    constexpr ssize_t elem_size = sizeof(typename props::Scalar);
+    array a;
+    if (props::vector) {
+        a = array({src.size()}, {elem_size * src.innerStride()}, src.data(), base);
+    } else {
+        a = array({src.rows(), src.cols()},
+                  {elem_size * src.rowStride(), elem_size * src.colStride()},
+                  src.data(),
+                  base);
+    }
+
+    if (!writeable) {
+        array_proxy(a.ptr())->flags &= ~detail::npy_api::NPY_ARRAY_WRITEABLE_;
+    }
+
+    return a.release();
+}
+
+// Takes an lvalue ref to some Eigen type and a (python) base object, creating a numpy array that
+// reference the Eigen object's data with `base` as the python-registered base class (if omitted,
+// the base will be set to None, and lifetime management is up to the caller).  The numpy array is
+// non-writeable if the given type is const.
+template <typename props, typename Type>
+handle eigen_ref_array(Type &src, handle parent = none()) {
+    // none here is to get past array's should-we-copy detection, which currently always
+    // copies when there is no base.  Setting the base to None should be harmless.
+    return eigen_array_cast<props>(src, parent, !std::is_const<Type>::value);
+}
+
+// Takes a pointer to some dense, plain Eigen type, builds a capsule around it, then returns a
+// numpy array that references the encapsulated data with a python-side reference to the capsule to
+// tie its destruction to that of any dependent python objects.  Const-ness is determined by
+// whether or not the Type of the pointer given is const.
+template <typename props, typename Type, typename = enable_if_t<is_eigen_dense_plain<Type>::value>>
+handle eigen_encapsulate(Type *src) {
+    capsule base(src, [](void *o) { delete static_cast<Type *>(o); });
+    return eigen_ref_array<props>(*src, base);
+}
+
+// Type caster for regular, dense matrix types (e.g. MatrixXd), but not maps/refs/etc. of dense
+// types.
+template <typename Type>
+struct type_caster<Type, enable_if_t<is_eigen_dense_plain<Type>::value>> {
+    using Scalar = typename Type::Scalar;
+    static_assert(!std::is_pointer<Scalar>::value,
+                  PYBIND11_EIGEN_MESSAGE_POINTER_TYPES_ARE_NOT_SUPPORTED);
+    using props = EigenProps<Type>;
+
+    bool load(handle src, bool convert) {
+        // If we're in no-convert mode, only load if given an array of the correct type
+        if (!convert && !isinstance<array_t<Scalar>>(src)) {
+            return false;
+        }
+
+        // Coerce into an array, but don't do type conversion yet; the copy below handles it.
+        auto buf = array::ensure(src);
+
+        if (!buf) {
+            return false;
+        }
+
+        auto dims = buf.ndim();
+        if (dims < 1 || dims > 2) {
+            return false;
+        }
+
+        auto fits = props::conformable(buf);
+        if (!fits) {
+            return false;
+        }
+
+        // Allocate the new type, then build a numpy reference into it
+        value = Type(fits.rows, fits.cols);
+        auto ref = reinterpret_steal<array>(eigen_ref_array<props>(value));
+        if (dims == 1) {
+            ref = ref.squeeze();
+        } else if (ref.ndim() == 1) {
+            buf = buf.squeeze();
+        }
+
+        int result = detail::npy_api::get().PyArray_CopyInto_(ref.ptr(), buf.ptr());
+
+        if (result < 0) { // Copy failed!
+            PyErr_Clear();
+            return false;
+        }
+
+        return true;
+    }
+
+private:
+    // Cast implementation
+    template <typename CType>
+    static handle cast_impl(CType *src, return_value_policy policy, handle parent) {
+        switch (policy) {
+            case return_value_policy::take_ownership:
+            case return_value_policy::automatic:
+                return eigen_encapsulate<props>(src);
+            case return_value_policy::move:
+                return eigen_encapsulate<props>(new CType(std::move(*src)));
+            case return_value_policy::copy:
+                return eigen_array_cast<props>(*src);
+            case return_value_policy::reference:
+            case return_value_policy::automatic_reference:
+                return eigen_ref_array<props>(*src);
+            case return_value_policy::reference_internal:
+                return eigen_ref_array<props>(*src, parent);
+            default:
+                throw cast_error("unhandled return_value_policy: should not happen!");
+        };
+    }
+
+public:
+    // Normal returned non-reference, non-const value:
+    static handle cast(Type &&src, return_value_policy /* policy */, handle parent) {
+        return cast_impl(&src, return_value_policy::move, parent);
+    }
+    // If you return a non-reference const, we mark the numpy array readonly:
+    static handle cast(const Type &&src, return_value_policy /* policy */, handle parent) {
+        return cast_impl(&src, return_value_policy::move, parent);
+    }
+    // lvalue reference return; default (automatic) becomes copy
+    static handle cast(Type &src, return_value_policy policy, handle parent) {
+        if (policy == return_value_policy::automatic
+            || policy == return_value_policy::automatic_reference) {
+            policy = return_value_policy::copy;
+        }
+        return cast_impl(&src, policy, parent);
+    }
+    // const lvalue reference return; default (automatic) becomes copy
+    static handle cast(const Type &src, return_value_policy policy, handle parent) {
+        if (policy == return_value_policy::automatic
+            || policy == return_value_policy::automatic_reference) {
+            policy = return_value_policy::copy;
+        }
+        return cast(&src, policy, parent);
+    }
+    // non-const pointer return
+    static handle cast(Type *src, return_value_policy policy, handle parent) {
+        return cast_impl(src, policy, parent);
+    }
+    // const pointer return
+    static handle cast(const Type *src, return_value_policy policy, handle parent) {
+        return cast_impl(src, policy, parent);
+    }
+
+    static constexpr auto name = props::descriptor;
+
+    // NOLINTNEXTLINE(google-explicit-constructor)
+    operator Type *() { return &value; }
+    // NOLINTNEXTLINE(google-explicit-constructor)
+    operator Type &() { return value; }
+    // NOLINTNEXTLINE(google-explicit-constructor)
+    operator Type &&() && { return std::move(value); }
+    template <typename T>
+    using cast_op_type = movable_cast_op_type<T>;
+
+private:
+    Type value;
+};
+
+// Base class for casting reference/map/block/etc. objects back to python.
+template <typename MapType>
+struct eigen_map_caster {
+    static_assert(!std::is_pointer<typename MapType::Scalar>::value,
+                  PYBIND11_EIGEN_MESSAGE_POINTER_TYPES_ARE_NOT_SUPPORTED);
+
+private:
+    using props = EigenProps<MapType>;
+
+public:
+    // Directly referencing a ref/map's data is a bit dangerous (whatever the map/ref points to has
+    // to stay around), but we'll allow it under the assumption that you know what you're doing
+    // (and have an appropriate keep_alive in place).  We return a numpy array pointing directly at
+    // the ref's data (The numpy array ends up read-only if the ref was to a const matrix type.)
+    // Note that this means you need to ensure you don't destroy the object in some other way (e.g.
+    // with an appropriate keep_alive, or with a reference to a statically allocated matrix).
+    static handle cast(const MapType &src, return_value_policy policy, handle parent) {
+        switch (policy) {
+            case return_value_policy::copy:
+                return eigen_array_cast<props>(src);
+            case return_value_policy::reference_internal:
+                return eigen_array_cast<props>(src, parent, is_eigen_mutable_map<MapType>::value);
+            case return_value_policy::reference:
+            case return_value_policy::automatic:
+            case return_value_policy::automatic_reference:
+                return eigen_array_cast<props>(src, none(), is_eigen_mutable_map<MapType>::value);
+            default:
+                // move, take_ownership don't make any sense for a ref/map:
+                pybind11_fail("Invalid return_value_policy for Eigen Map/Ref/Block type");
+        }
+    }
+
+    static constexpr auto name = props::descriptor;
+
+    // Explicitly delete these: support python -> C++ conversion on these (i.e. these can be return
+    // types but not bound arguments).  We still provide them (with an explicitly delete) so that
+    // you end up here if you try anyway.
+    bool load(handle, bool) = delete;
+    operator MapType() = delete;
+    template <typename>
+    using cast_op_type = MapType;
+};
+
+// We can return any map-like object (but can only load Refs, specialized next):
+template <typename Type>
+struct type_caster<Type, enable_if_t<is_eigen_dense_map<Type>::value>> : eigen_map_caster<Type> {};
+
+// Loader for Ref<...> arguments.  See the documentation for info on how to make this work without
+// copying (it requires some extra effort in many cases).
+template <typename PlainObjectType, typename StrideType>
+struct type_caster<
+    Eigen::Ref<PlainObjectType, 0, StrideType>,
+    enable_if_t<is_eigen_dense_map<Eigen::Ref<PlainObjectType, 0, StrideType>>::value>>
+    : public eigen_map_caster<Eigen::Ref<PlainObjectType, 0, StrideType>> {
+private:
+    using Type = Eigen::Ref<PlainObjectType, 0, StrideType>;
+    using props = EigenProps<Type>;
+    using Scalar = typename props::Scalar;
+    static_assert(!std::is_pointer<Scalar>::value,
+                  PYBIND11_EIGEN_MESSAGE_POINTER_TYPES_ARE_NOT_SUPPORTED);
+    using MapType = Eigen::Map<PlainObjectType, 0, StrideType>;
+    using Array
+        = array_t<Scalar,
+                  array::forcecast
+                      | ((props::row_major ? props::inner_stride : props::outer_stride) == 1
+                             ? array::c_style
+                         : (props::row_major ? props::outer_stride : props::inner_stride) == 1
+                             ? array::f_style
+                             : 0)>;
+    static constexpr bool need_writeable = is_eigen_mutable_map<Type>::value;
+    // Delay construction (these have no default constructor)
+    std::unique_ptr<MapType> map;
+    std::unique_ptr<Type> ref;
+    // Our array.  When possible, this is just a numpy array pointing to the source data, but
+    // sometimes we can't avoid copying (e.g. input is not a numpy array at all, has an
+    // incompatible layout, or is an array of a type that needs to be converted).  Using a numpy
+    // temporary (rather than an Eigen temporary) saves an extra copy when we need both type
+    // conversion and storage order conversion.  (Note that we refuse to use this temporary copy
+    // when loading an argument for a Ref<M> with M non-const, i.e. a read-write reference).
+    Array copy_or_ref;
+
+public:
+    bool load(handle src, bool convert) {
+        // First check whether what we have is already an array of the right type.  If not, we
+        // can't avoid a copy (because the copy is also going to do type conversion).
+        bool need_copy = !isinstance<Array>(src);
+
+        EigenConformable<props::row_major> fits;
+        if (!need_copy) {
+            // We don't need a converting copy, but we also need to check whether the strides are
+            // compatible with the Ref's stride requirements
+            auto aref = reinterpret_borrow<Array>(src);
+
+            if (aref && (!need_writeable || aref.writeable())) {
+                fits = props::conformable(aref);
+                if (!fits) {
+                    return false; // Incompatible dimensions
+                }
+                if (!fits.template stride_compatible<props>()) {
+                    need_copy = true;
+                } else {
+                    copy_or_ref = std::move(aref);
+                }
+            } else {
+                need_copy = true;
+            }
+        }
+
+        if (need_copy) {
+            // We need to copy: If we need a mutable reference, or we're not supposed to convert
+            // (either because we're in the no-convert overload pass, or because we're explicitly
+            // instructed not to copy (via `py::arg().noconvert()`) we have to fail loading.
+            if (!convert || need_writeable) {
+                return false;
+            }
+
+            Array copy = Array::ensure(src);
+            if (!copy) {
+                return false;
+            }
+            fits = props::conformable(copy);
+            if (!fits || !fits.template stride_compatible<props>()) {
+                return false;
+            }
+            copy_or_ref = std::move(copy);
+            loader_life_support::add_patient(copy_or_ref);
+        }
+
+        ref.reset();
+        map.reset(new MapType(data(copy_or_ref),
+                              fits.rows,
+                              fits.cols,
+                              make_stride(fits.stride.outer(), fits.stride.inner())));
+        ref.reset(new Type(*map));
+
+        return true;
+    }
+
+    // NOLINTNEXTLINE(google-explicit-constructor)
+    operator Type *() { return ref.get(); }
+    // NOLINTNEXTLINE(google-explicit-constructor)
+    operator Type &() { return *ref; }
+    template <typename _T>
+    using cast_op_type = pybind11::detail::cast_op_type<_T>;
+
+private:
+    template <typename T = Type, enable_if_t<is_eigen_mutable_map<T>::value, int> = 0>
+    Scalar *data(Array &a) {
+        return a.mutable_data();
+    }
+
+    template <typename T = Type, enable_if_t<!is_eigen_mutable_map<T>::value, int> = 0>
+    const Scalar *data(Array &a) {
+        return a.data();
+    }
+
+    // Attempt to figure out a constructor of `Stride` that will work.
+    // If both strides are fixed, use a default constructor:
+    template <typename S>
+    using stride_ctor_default = bool_constant<S::InnerStrideAtCompileTime != Eigen::Dynamic
+                                              && S::OuterStrideAtCompileTime != Eigen::Dynamic
+                                              && std::is_default_constructible<S>::value>;
+    // Otherwise, if there is a two-index constructor, assume it is (outer,inner) like
+    // Eigen::Stride, and use it:
+    template <typename S>
+    using stride_ctor_dual
+        = bool_constant<!stride_ctor_default<S>::value
+                        && std::is_constructible<S, EigenIndex, EigenIndex>::value>;
+    // Otherwise, if there is a one-index constructor, and just one of the strides is dynamic, use
+    // it (passing whichever stride is dynamic).
+    template <typename S>
+    using stride_ctor_outer
+        = bool_constant<!any_of<stride_ctor_default<S>, stride_ctor_dual<S>>::value
+                        && S::OuterStrideAtCompileTime == Eigen::Dynamic
+                        && S::InnerStrideAtCompileTime != Eigen::Dynamic
+                        && std::is_constructible<S, EigenIndex>::value>;
+    template <typename S>
+    using stride_ctor_inner
+        = bool_constant<!any_of<stride_ctor_default<S>, stride_ctor_dual<S>>::value
+                        && S::InnerStrideAtCompileTime == Eigen::Dynamic
+                        && S::OuterStrideAtCompileTime != Eigen::Dynamic
+                        && std::is_constructible<S, EigenIndex>::value>;
+
+    template <typename S = StrideType, enable_if_t<stride_ctor_default<S>::value, int> = 0>
+    static S make_stride(EigenIndex, EigenIndex) {
+        return S();
+    }
+    template <typename S = StrideType, enable_if_t<stride_ctor_dual<S>::value, int> = 0>
+    static S make_stride(EigenIndex outer, EigenIndex inner) {
+        return S(outer, inner);
+    }
+    template <typename S = StrideType, enable_if_t<stride_ctor_outer<S>::value, int> = 0>
+    static S make_stride(EigenIndex outer, EigenIndex) {
+        return S(outer);
+    }
+    template <typename S = StrideType, enable_if_t<stride_ctor_inner<S>::value, int> = 0>
+    static S make_stride(EigenIndex, EigenIndex inner) {
+        return S(inner);
+    }
+};
+
+// type_caster for special matrix types (e.g. DiagonalMatrix), which are EigenBase, but not
+// EigenDense (i.e. they don't have a data(), at least not with the usual matrix layout).
+// load() is not supported, but we can cast them into the python domain by first copying to a
+// regular Eigen::Matrix, then casting that.
+template <typename Type>
+struct type_caster<Type, enable_if_t<is_eigen_other<Type>::value>> {
+    static_assert(!std::is_pointer<typename Type::Scalar>::value,
+                  PYBIND11_EIGEN_MESSAGE_POINTER_TYPES_ARE_NOT_SUPPORTED);
+
+protected:
+    using Matrix
+        = Eigen::Matrix<typename Type::Scalar, Type::RowsAtCompileTime, Type::ColsAtCompileTime>;
+    using props = EigenProps<Matrix>;
+
+public:
+    static handle cast(const Type &src, return_value_policy /* policy */, handle /* parent */) {
+        handle h = eigen_encapsulate<props>(new Matrix(src));
+        return h;
+    }
+    static handle cast(const Type *src, return_value_policy policy, handle parent) {
+        return cast(*src, policy, parent);
+    }
+
+    static constexpr auto name = props::descriptor;
+
+    // Explicitly delete these: support python -> C++ conversion on these (i.e. these can be return
+    // types but not bound arguments).  We still provide them (with an explicitly delete) so that
+    // you end up here if you try anyway.
+    bool load(handle, bool) = delete;
+    operator Type() = delete;
+    template <typename>
+    using cast_op_type = Type;
+};
+
+template <typename Type>
+struct type_caster<Type, enable_if_t<is_eigen_sparse<Type>::value>> {
+    using Scalar = typename Type::Scalar;
+    static_assert(!std::is_pointer<Scalar>::value,
+                  PYBIND11_EIGEN_MESSAGE_POINTER_TYPES_ARE_NOT_SUPPORTED);
+    using StorageIndex = remove_reference_t<decltype(*std::declval<Type>().outerIndexPtr())>;
+    using Index = typename Type::Index;
+    static constexpr bool rowMajor = Type::IsRowMajor;
+
+    bool load(handle src, bool) {
+        if (!src) {
+            return false;
+        }
+
+        auto obj = reinterpret_borrow<object>(src);
+        object sparse_module = module_::import("scipy.sparse");
+        object matrix_type = sparse_module.attr(rowMajor ? "csr_matrix" : "csc_matrix");
+
+        if (!type::handle_of(obj).is(matrix_type)) {
+            try {
+                obj = matrix_type(obj);
+            } catch (const error_already_set &) {
+                return false;
+            }
+        }
+
+        auto values = array_t<Scalar>((object) obj.attr("data"));
+        auto innerIndices = array_t<StorageIndex>((object) obj.attr("indices"));
+        auto outerIndices = array_t<StorageIndex>((object) obj.attr("indptr"));
+        auto shape = pybind11::tuple((pybind11::object) obj.attr("shape"));
+        auto nnz = obj.attr("nnz").cast<Index>();
+
+        if (!values || !innerIndices || !outerIndices) {
+            return false;
+        }
+
+        value = EigenMapSparseMatrix<Scalar,
+                                     Type::Flags &(Eigen::RowMajor | Eigen::ColMajor),
+                                     StorageIndex>(shape[0].cast<Index>(),
+                                                   shape[1].cast<Index>(),
+                                                   std::move(nnz),
+                                                   outerIndices.mutable_data(),
+                                                   innerIndices.mutable_data(),
+                                                   values.mutable_data());
+
+        return true;
+    }
+
+    static handle cast(const Type &src, return_value_policy /* policy */, handle /* parent */) {
+        const_cast<Type &>(src).makeCompressed();
+
+        object matrix_type
+            = module_::import("scipy.sparse").attr(rowMajor ? "csr_matrix" : "csc_matrix");
+
+        array data(src.nonZeros(), src.valuePtr());
+        array outerIndices((rowMajor ? src.rows() : src.cols()) + 1, src.outerIndexPtr());
+        array innerIndices(src.nonZeros(), src.innerIndexPtr());
+
+        return matrix_type(pybind11::make_tuple(
+                               std::move(data), std::move(innerIndices), std::move(outerIndices)),
+                           pybind11::make_tuple(src.rows(), src.cols()))
+            .release();
+    }
+
+    PYBIND11_TYPE_CASTER(Type,
+                         const_name<(Type::IsRowMajor) != 0>("scipy.sparse.csr_matrix[",
+                                                             "scipy.sparse.csc_matrix[")
+                             + npy_format_descriptor<Scalar>::name + const_name("]"));
+};
+
+PYBIND11_NAMESPACE_END(detail)
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/eigen/tensor.h

@@ -0,0 +1,515 @@
+/*
+    pybind11/eigen/tensor.h: Transparent conversion for Eigen tensors
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include <pybind11/numpy.h>
+
+#include "common.h"
+
+#if defined(__GNUC__) && !defined(__clang__) && !defined(__INTEL_COMPILER)
+static_assert(__GNUC__ > 5, "Eigen Tensor support in pybind11 requires GCC > 5.0");
+#endif
+
+// Disable warnings for Eigen
+PYBIND11_WARNING_PUSH
+PYBIND11_WARNING_DISABLE_MSVC(4554)
+PYBIND11_WARNING_DISABLE_MSVC(4127)
+#if defined(__MINGW32__)
+PYBIND11_WARNING_DISABLE_GCC("-Wmaybe-uninitialized")
+#endif
+
+#include <unsupported/Eigen/CXX11/Tensor>
+
+PYBIND11_WARNING_POP
+
+static_assert(EIGEN_VERSION_AT_LEAST(3, 3, 0),
+              "Eigen Tensor support in pybind11 requires Eigen >= 3.3.0");
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+PYBIND11_WARNING_DISABLE_MSVC(4127)
+
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+inline bool is_tensor_aligned(const void *data) {
+    return (reinterpret_cast<std::size_t>(data) % EIGEN_DEFAULT_ALIGN_BYTES) == 0;
+}
+
+template <typename T>
+constexpr int compute_array_flag_from_tensor() {
+    static_assert((static_cast<int>(T::Layout) == static_cast<int>(Eigen::RowMajor))
+                      || (static_cast<int>(T::Layout) == static_cast<int>(Eigen::ColMajor)),
+                  "Layout must be row or column major");
+    return (static_cast<int>(T::Layout) == static_cast<int>(Eigen::RowMajor)) ? array::c_style
+                                                                              : array::f_style;
+}
+
+template <typename T>
+struct eigen_tensor_helper {};
+
+template <typename Scalar_, int NumIndices_, int Options_, typename IndexType>
+struct eigen_tensor_helper<Eigen::Tensor<Scalar_, NumIndices_, Options_, IndexType>> {
+    using Type = Eigen::Tensor<Scalar_, NumIndices_, Options_, IndexType>;
+    using ValidType = void;
+
+    static Eigen::DSizes<typename Type::Index, Type::NumIndices> get_shape(const Type &f) {
+        return f.dimensions();
+    }
+
+    static constexpr bool
+    is_correct_shape(const Eigen::DSizes<typename Type::Index, Type::NumIndices> & /*shape*/) {
+        return true;
+    }
+
+    template <typename T>
+    struct helper {};
+
+    template <size_t... Is>
+    struct helper<index_sequence<Is...>> {
+        static constexpr auto value = ::pybind11::detail::concat(const_name(((void) Is, "?"))...);
+    };
+
+    static constexpr auto dimensions_descriptor
+        = helper<decltype(make_index_sequence<Type::NumIndices>())>::value;
+
+    template <typename... Args>
+    static Type *alloc(Args &&...args) {
+        return new Type(std::forward<Args>(args)...);
+    }
+
+    static void free(Type *tensor) { delete tensor; }
+};
+
+template <typename Scalar_, typename std::ptrdiff_t... Indices, int Options_, typename IndexType>
+struct eigen_tensor_helper<
+    Eigen::TensorFixedSize<Scalar_, Eigen::Sizes<Indices...>, Options_, IndexType>> {
+    using Type = Eigen::TensorFixedSize<Scalar_, Eigen::Sizes<Indices...>, Options_, IndexType>;
+    using ValidType = void;
+
+    static constexpr Eigen::DSizes<typename Type::Index, Type::NumIndices>
+    get_shape(const Type & /*f*/) {
+        return get_shape();
+    }
+
+    static constexpr Eigen::DSizes<typename Type::Index, Type::NumIndices> get_shape() {
+        return Eigen::DSizes<typename Type::Index, Type::NumIndices>(Indices...);
+    }
+
+    static bool
+    is_correct_shape(const Eigen::DSizes<typename Type::Index, Type::NumIndices> &shape) {
+        return get_shape() == shape;
+    }
+
+    static constexpr auto dimensions_descriptor
+        = ::pybind11::detail::concat(const_name<Indices>()...);
+
+    template <typename... Args>
+    static Type *alloc(Args &&...args) {
+        Eigen::aligned_allocator<Type> allocator;
+        return ::new (allocator.allocate(1)) Type(std::forward<Args>(args)...);
+    }
+
+    static void free(Type *tensor) {
+        Eigen::aligned_allocator<Type> allocator;
+        tensor->~Type();
+        allocator.deallocate(tensor, 1);
+    }
+};
+
+template <typename Type, bool ShowDetails, bool NeedsWriteable = false>
+struct get_tensor_descriptor {
+    static constexpr auto details
+        = const_name<NeedsWriteable>(", flags.writeable", "")
+          + const_name<static_cast<int>(Type::Layout) == static_cast<int>(Eigen::RowMajor)>(
+              ", flags.c_contiguous", ", flags.f_contiguous");
+    static constexpr auto value
+        = const_name("numpy.ndarray[") + npy_format_descriptor<typename Type::Scalar>::name
+          + const_name("[") + eigen_tensor_helper<remove_cv_t<Type>>::dimensions_descriptor
+          + const_name("]") + const_name<ShowDetails>(details, const_name("")) + const_name("]");
+};
+
+// When EIGEN_AVOID_STL_ARRAY is defined, Eigen::DSizes<T, 0> does not have the begin() member
+// function. Falling back to a simple loop works around this issue.
+//
+// We need to disable the type-limits warning for the inner loop when size = 0.
+
+PYBIND11_WARNING_PUSH
+PYBIND11_WARNING_DISABLE_GCC("-Wtype-limits")
+
+template <typename T, int size>
+std::vector<T> convert_dsizes_to_vector(const Eigen::DSizes<T, size> &arr) {
+    std::vector<T> result(size);
+
+    for (size_t i = 0; i < size; i++) {
+        result[i] = arr[i];
+    }
+
+    return result;
+}
+
+template <typename T, int size>
+Eigen::DSizes<T, size> get_shape_for_array(const array &arr) {
+    Eigen::DSizes<T, size> result;
+    const T *shape = arr.shape();
+    for (size_t i = 0; i < size; i++) {
+        result[i] = shape[i];
+    }
+
+    return result;
+}
+
+PYBIND11_WARNING_POP
+
+template <typename Type>
+struct type_caster<Type, typename eigen_tensor_helper<Type>::ValidType> {
+    static_assert(!std::is_pointer<typename Type::Scalar>::value,
+                  PYBIND11_EIGEN_MESSAGE_POINTER_TYPES_ARE_NOT_SUPPORTED);
+    using Helper = eigen_tensor_helper<Type>;
+    static constexpr auto temp_name = get_tensor_descriptor<Type, false>::value;
+    PYBIND11_TYPE_CASTER(Type, temp_name);
+
+    bool load(handle src, bool convert) {
+        if (!convert) {
+            if (!isinstance<array>(src)) {
+                return false;
+            }
+            array temp = array::ensure(src);
+            if (!temp) {
+                return false;
+            }
+
+            if (!temp.dtype().is(dtype::of<typename Type::Scalar>())) {
+                return false;
+            }
+        }
+
+        array_t<typename Type::Scalar, compute_array_flag_from_tensor<Type>()> arr(
+            reinterpret_borrow<object>(src));
+
+        if (arr.ndim() != Type::NumIndices) {
+            return false;
+        }
+        auto shape = get_shape_for_array<typename Type::Index, Type::NumIndices>(arr);
+
+        if (!Helper::is_correct_shape(shape)) {
+            return false;
+        }
+
+#if EIGEN_VERSION_AT_LEAST(3, 4, 0)
+        auto data_pointer = arr.data();
+#else
+        // Handle Eigen bug
+        auto data_pointer = const_cast<typename Type::Scalar *>(arr.data());
+#endif
+
+        if (is_tensor_aligned(arr.data())) {
+            value = Eigen::TensorMap<const Type, Eigen::Aligned>(data_pointer, shape);
+        } else {
+            value = Eigen::TensorMap<const Type>(data_pointer, shape);
+        }
+
+        return true;
+    }
+
+    static handle cast(Type &&src, return_value_policy policy, handle parent) {
+        if (policy == return_value_policy::reference
+            || policy == return_value_policy::reference_internal) {
+            pybind11_fail("Cannot use a reference return value policy for an rvalue");
+        }
+        return cast_impl(&src, return_value_policy::move, parent);
+    }
+
+    static handle cast(const Type &&src, return_value_policy policy, handle parent) {
+        if (policy == return_value_policy::reference
+            || policy == return_value_policy::reference_internal) {
+            pybind11_fail("Cannot use a reference return value policy for an rvalue");
+        }
+        return cast_impl(&src, return_value_policy::move, parent);
+    }
+
+    static handle cast(Type &src, return_value_policy policy, handle parent) {
+        if (policy == return_value_policy::automatic
+            || policy == return_value_policy::automatic_reference) {
+            policy = return_value_policy::copy;
+        }
+        return cast_impl(&src, policy, parent);
+    }
+
+    static handle cast(const Type &src, return_value_policy policy, handle parent) {
+        if (policy == return_value_policy::automatic
+            || policy == return_value_policy::automatic_reference) {
+            policy = return_value_policy::copy;
+        }
+        return cast(&src, policy, parent);
+    }
+
+    static handle cast(Type *src, return_value_policy policy, handle parent) {
+        if (policy == return_value_policy::automatic) {
+            policy = return_value_policy::take_ownership;
+        } else if (policy == return_value_policy::automatic_reference) {
+            policy = return_value_policy::reference;
+        }
+        return cast_impl(src, policy, parent);
+    }
+
+    static handle cast(const Type *src, return_value_policy policy, handle parent) {
+        if (policy == return_value_policy::automatic) {
+            policy = return_value_policy::take_ownership;
+        } else if (policy == return_value_policy::automatic_reference) {
+            policy = return_value_policy::reference;
+        }
+        return cast_impl(src, policy, parent);
+    }
+
+    template <typename C>
+    static handle cast_impl(C *src, return_value_policy policy, handle parent) {
+        object parent_object;
+        bool writeable = false;
+        switch (policy) {
+            case return_value_policy::move:
+                if (std::is_const<C>::value) {
+                    pybind11_fail("Cannot move from a constant reference");
+                }
+
+                src = Helper::alloc(std::move(*src));
+
+                parent_object
+                    = capsule(src, [](void *ptr) { Helper::free(reinterpret_cast<Type *>(ptr)); });
+                writeable = true;
+                break;
+
+            case return_value_policy::take_ownership:
+                if (std::is_const<C>::value) {
+                    // This cast is ugly, and might be UB in some cases, but we don't have an
+                    // alternative here as we must free that memory
+                    Helper::free(const_cast<Type *>(src));
+                    pybind11_fail("Cannot take ownership of a const reference");
+                }
+
+                parent_object
+                    = capsule(src, [](void *ptr) { Helper::free(reinterpret_cast<Type *>(ptr)); });
+                writeable = true;
+                break;
+
+            case return_value_policy::copy:
+                writeable = true;
+                break;
+
+            case return_value_policy::reference:
+                parent_object = none();
+                writeable = !std::is_const<C>::value;
+                break;
+
+            case return_value_policy::reference_internal:
+                // Default should do the right thing
+                if (!parent) {
+                    pybind11_fail("Cannot use reference internal when there is no parent");
+                }
+                parent_object = reinterpret_borrow<object>(parent);
+                writeable = !std::is_const<C>::value;
+                break;
+
+            default:
+                pybind11_fail("pybind11 bug in eigen.h, please file a bug report");
+        }
+
+        auto result = array_t<typename Type::Scalar, compute_array_flag_from_tensor<Type>()>(
+            convert_dsizes_to_vector(Helper::get_shape(*src)), src->data(), parent_object);
+
+        if (!writeable) {
+            array_proxy(result.ptr())->flags &= ~detail::npy_api::NPY_ARRAY_WRITEABLE_;
+        }
+
+        return result.release();
+    }
+};
+
+template <typename StoragePointerType,
+          bool needs_writeable,
+          enable_if_t<!needs_writeable, bool> = true>
+StoragePointerType get_array_data_for_type(array &arr) {
+#if EIGEN_VERSION_AT_LEAST(3, 4, 0)
+    return reinterpret_cast<StoragePointerType>(arr.data());
+#else
+    // Handle Eigen bug
+    return reinterpret_cast<StoragePointerType>(const_cast<void *>(arr.data()));
+#endif
+}
+
+template <typename StoragePointerType,
+          bool needs_writeable,
+          enable_if_t<needs_writeable, bool> = true>
+StoragePointerType get_array_data_for_type(array &arr) {
+    return reinterpret_cast<StoragePointerType>(arr.mutable_data());
+}
+
+template <typename T, typename = void>
+struct get_storage_pointer_type;
+
+template <typename MapType>
+struct get_storage_pointer_type<MapType, void_t<typename MapType::StoragePointerType>> {
+    using SPT = typename MapType::StoragePointerType;
+};
+
+template <typename MapType>
+struct get_storage_pointer_type<MapType, void_t<typename MapType::PointerArgType>> {
+    using SPT = typename MapType::PointerArgType;
+};
+
+template <typename Type, int Options>
+struct type_caster<Eigen::TensorMap<Type, Options>,
+                   typename eigen_tensor_helper<remove_cv_t<Type>>::ValidType> {
+    static_assert(!std::is_pointer<typename Type::Scalar>::value,
+                  PYBIND11_EIGEN_MESSAGE_POINTER_TYPES_ARE_NOT_SUPPORTED);
+    using MapType = Eigen::TensorMap<Type, Options>;
+    using Helper = eigen_tensor_helper<remove_cv_t<Type>>;
+
+    bool load(handle src, bool /*convert*/) {
+        // Note that we have a lot more checks here as we want to make sure to avoid copies
+        if (!isinstance<array>(src)) {
+            return false;
+        }
+        auto arr = reinterpret_borrow<array>(src);
+        if ((arr.flags() & compute_array_flag_from_tensor<Type>()) == 0) {
+            return false;
+        }
+
+        if (!arr.dtype().is(dtype::of<typename Type::Scalar>())) {
+            return false;
+        }
+
+        if (arr.ndim() != Type::NumIndices) {
+            return false;
+        }
+
+        constexpr bool is_aligned = (Options & Eigen::Aligned) != 0;
+
+        if (is_aligned && !is_tensor_aligned(arr.data())) {
+            return false;
+        }
+
+        auto shape = get_shape_for_array<typename Type::Index, Type::NumIndices>(arr);
+
+        if (!Helper::is_correct_shape(shape)) {
+            return false;
+        }
+
+        if (needs_writeable && !arr.writeable()) {
+            return false;
+        }
+
+        auto result = get_array_data_for_type<typename get_storage_pointer_type<MapType>::SPT,
+                                              needs_writeable>(arr);
+
+        value.reset(new MapType(std::move(result), std::move(shape)));
+
+        return true;
+    }
+
+    static handle cast(MapType &&src, return_value_policy policy, handle parent) {
+        return cast_impl(&src, policy, parent);
+    }
+
+    static handle cast(const MapType &&src, return_value_policy policy, handle parent) {
+        return cast_impl(&src, policy, parent);
+    }
+
+    static handle cast(MapType &src, return_value_policy policy, handle parent) {
+        if (policy == return_value_policy::automatic
+            || policy == return_value_policy::automatic_reference) {
+            policy = return_value_policy::copy;
+        }
+        return cast_impl(&src, policy, parent);
+    }
+
+    static handle cast(const MapType &src, return_value_policy policy, handle parent) {
+        if (policy == return_value_policy::automatic
+            || policy == return_value_policy::automatic_reference) {
+            policy = return_value_policy::copy;
+        }
+        return cast(&src, policy, parent);
+    }
+
+    static handle cast(MapType *src, return_value_policy policy, handle parent) {
+        if (policy == return_value_policy::automatic) {
+            policy = return_value_policy::take_ownership;
+        } else if (policy == return_value_policy::automatic_reference) {
+            policy = return_value_policy::reference;
+        }
+        return cast_impl(src, policy, parent);
+    }
+
+    static handle cast(const MapType *src, return_value_policy policy, handle parent) {
+        if (policy == return_value_policy::automatic) {
+            policy = return_value_policy::take_ownership;
+        } else if (policy == return_value_policy::automatic_reference) {
+            policy = return_value_policy::reference;
+        }
+        return cast_impl(src, policy, parent);
+    }
+
+    template <typename C>
+    static handle cast_impl(C *src, return_value_policy policy, handle parent) {
+        object parent_object;
+        constexpr bool writeable = !std::is_const<C>::value;
+        switch (policy) {
+            case return_value_policy::reference:
+                parent_object = none();
+                break;
+
+            case return_value_policy::reference_internal:
+                // Default should do the right thing
+                if (!parent) {
+                    pybind11_fail("Cannot use reference internal when there is no parent");
+                }
+                parent_object = reinterpret_borrow<object>(parent);
+                break;
+
+            default:
+                // move, take_ownership don't make any sense for a ref/map:
+                pybind11_fail("Invalid return_value_policy for Eigen Map type, must be either "
+                              "reference or reference_internal");
+        }
+
+        auto result = array_t<typename Type::Scalar, compute_array_flag_from_tensor<Type>()>(
+            convert_dsizes_to_vector(Helper::get_shape(*src)),
+            src->data(),
+            std::move(parent_object));
+
+        if (!writeable) {
+            array_proxy(result.ptr())->flags &= ~detail::npy_api::NPY_ARRAY_WRITEABLE_;
+        }
+
+        return result.release();
+    }
+
+#if EIGEN_VERSION_AT_LEAST(3, 4, 0)
+
+    static constexpr bool needs_writeable = !std::is_const<typename std::remove_pointer<
+        typename get_storage_pointer_type<MapType>::SPT>::type>::value;
+#else
+    // Handle Eigen bug
+    static constexpr bool needs_writeable = !std::is_const<Type>::value;
+#endif
+
+protected:
+    // TODO: Move to std::optional once std::optional has more support
+    std::unique_ptr<MapType> value;
+
+public:
+    static constexpr auto name = get_tensor_descriptor<Type, true, needs_writeable>::value;
+    explicit operator MapType *() { return value.get(); }
+    explicit operator MapType &() { return *value; }
+    explicit operator MapType &&() && { return std::move(*value); }
+
+    template <typename T_>
+    using cast_op_type = ::pybind11::detail::movable_cast_op_type<T_>;
+};
+
+PYBIND11_NAMESPACE_END(detail)
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/embed.h

@@ -0,0 +1,313 @@
+/*
+    pybind11/embed.h: Support for embedding the interpreter
+
+    Copyright (c) 2017 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "pybind11.h"
+#include "eval.h"
+
+#include <memory>
+#include <vector>
+
+#if defined(PYPY_VERSION)
+#    error Embedding the interpreter is not supported with PyPy
+#endif
+
+#define PYBIND11_EMBEDDED_MODULE_IMPL(name)                                                       \
+    extern "C" PyObject *pybind11_init_impl_##name();                                             \
+    extern "C" PyObject *pybind11_init_impl_##name() { return pybind11_init_wrapper_##name(); }
+
+/** \rst
+    Add a new module to the table of builtins for the interpreter. Must be
+    defined in global scope. The first macro parameter is the name of the
+    module (without quotes). The second parameter is the variable which will
+    be used as the interface to add functions and classes to the module.
+
+    .. code-block:: cpp
+
+        PYBIND11_EMBEDDED_MODULE(example, m) {
+            // ... initialize functions and classes here
+            m.def("foo", []() {
+                return "Hello, World!";
+            });
+        }
+ \endrst */
+#define PYBIND11_EMBEDDED_MODULE(name, variable)                                                  \
+    static ::pybind11::module_::module_def PYBIND11_CONCAT(pybind11_module_def_, name);           \
+    static void PYBIND11_CONCAT(pybind11_init_, name)(::pybind11::module_ &);                     \
+    static PyObject PYBIND11_CONCAT(*pybind11_init_wrapper_, name)() {                            \
+        auto m = ::pybind11::module_::create_extension_module(                                    \
+            PYBIND11_TOSTRING(name), nullptr, &PYBIND11_CONCAT(pybind11_module_def_, name));      \
+        try {                                                                                     \
+            PYBIND11_CONCAT(pybind11_init_, name)(m);                                             \
+            return m.ptr();                                                                       \
+        }                                                                                         \
+        PYBIND11_CATCH_INIT_EXCEPTIONS                                                            \
+    }                                                                                             \
+    PYBIND11_EMBEDDED_MODULE_IMPL(name)                                                           \
+    ::pybind11::detail::embedded_module PYBIND11_CONCAT(pybind11_module_, name)(                  \
+        PYBIND11_TOSTRING(name), PYBIND11_CONCAT(pybind11_init_impl_, name));                     \
+    void PYBIND11_CONCAT(pybind11_init_, name)(::pybind11::module_                                \
+                                               & variable) // NOLINT(bugprone-macro-parentheses)
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+/// Python 2.7/3.x compatible version of `PyImport_AppendInittab` and error checks.
+struct embedded_module {
+    using init_t = PyObject *(*) ();
+    embedded_module(const char *name, init_t init) {
+        if (Py_IsInitialized() != 0) {
+            pybind11_fail("Can't add new modules after the interpreter has been initialized");
+        }
+
+        auto result = PyImport_AppendInittab(name, init);
+        if (result == -1) {
+            pybind11_fail("Insufficient memory to add a new module");
+        }
+    }
+};
+
+struct wide_char_arg_deleter {
+    void operator()(wchar_t *ptr) const {
+        // API docs: https://docs.python.org/3/c-api/sys.html#c.Py_DecodeLocale
+        PyMem_RawFree(ptr);
+    }
+};
+
+inline wchar_t *widen_chars(const char *safe_arg) {
+    wchar_t *widened_arg = Py_DecodeLocale(safe_arg, nullptr);
+    return widened_arg;
+}
+
+inline void precheck_interpreter() {
+    if (Py_IsInitialized() != 0) {
+        pybind11_fail("The interpreter is already running");
+    }
+}
+
+#if !defined(PYBIND11_PYCONFIG_SUPPORT_PY_VERSION_HEX)
+#    define PYBIND11_PYCONFIG_SUPPORT_PY_VERSION_HEX (0x03080000)
+#endif
+
+#if PY_VERSION_HEX < PYBIND11_PYCONFIG_SUPPORT_PY_VERSION_HEX
+inline void initialize_interpreter_pre_pyconfig(bool init_signal_handlers,
+                                                int argc,
+                                                const char *const *argv,
+                                                bool add_program_dir_to_path) {
+    detail::precheck_interpreter();
+    Py_InitializeEx(init_signal_handlers ? 1 : 0);
+
+    // Before it was special-cased in python 3.8, passing an empty or null argv
+    // caused a segfault, so we have to reimplement the special case ourselves.
+    bool special_case = (argv == nullptr || argc <= 0);
+
+    const char *const empty_argv[]{"\0"};
+    const char *const *safe_argv = special_case ? empty_argv : argv;
+    if (special_case) {
+        argc = 1;
+    }
+
+    auto argv_size = static_cast<size_t>(argc);
+    // SetArgv* on python 3 takes wchar_t, so we have to convert.
+    std::unique_ptr<wchar_t *[]> widened_argv(new wchar_t *[argv_size]);
+    std::vector<std::unique_ptr<wchar_t[], detail::wide_char_arg_deleter>> widened_argv_entries;
+    widened_argv_entries.reserve(argv_size);
+    for (size_t ii = 0; ii < argv_size; ++ii) {
+        widened_argv_entries.emplace_back(detail::widen_chars(safe_argv[ii]));
+        if (!widened_argv_entries.back()) {
+            // A null here indicates a character-encoding failure or the python
+            // interpreter out of memory. Give up.
+            return;
+        }
+        widened_argv[ii] = widened_argv_entries.back().get();
+    }
+
+    auto *pysys_argv = widened_argv.get();
+
+    PySys_SetArgvEx(argc, pysys_argv, static_cast<int>(add_program_dir_to_path));
+}
+#endif
+
+PYBIND11_NAMESPACE_END(detail)
+
+#if PY_VERSION_HEX >= PYBIND11_PYCONFIG_SUPPORT_PY_VERSION_HEX
+inline void initialize_interpreter(PyConfig *config,
+                                   int argc = 0,
+                                   const char *const *argv = nullptr,
+                                   bool add_program_dir_to_path = true) {
+    detail::precheck_interpreter();
+    PyStatus status = PyConfig_SetBytesArgv(config, argc, const_cast<char *const *>(argv));
+    if (PyStatus_Exception(status) != 0) {
+        // A failure here indicates a character-encoding failure or the python
+        // interpreter out of memory. Give up.
+        PyConfig_Clear(config);
+        throw std::runtime_error(PyStatus_IsError(status) != 0 ? status.err_msg
+                                                               : "Failed to prepare CPython");
+    }
+    status = Py_InitializeFromConfig(config);
+    if (PyStatus_Exception(status) != 0) {
+        PyConfig_Clear(config);
+        throw std::runtime_error(PyStatus_IsError(status) != 0 ? status.err_msg
+                                                               : "Failed to init CPython");
+    }
+    if (add_program_dir_to_path) {
+        PyRun_SimpleString("import sys, os.path; "
+                           "sys.path.insert(0, "
+                           "os.path.abspath(os.path.dirname(sys.argv[0])) "
+                           "if sys.argv and os.path.exists(sys.argv[0]) else '')");
+    }
+    PyConfig_Clear(config);
+}
+#endif
+
+/** \rst
+    Initialize the Python interpreter. No other pybind11 or CPython API functions can be
+    called before this is done; with the exception of `PYBIND11_EMBEDDED_MODULE`. The
+    optional `init_signal_handlers` parameter can be used to skip the registration of
+    signal handlers (see the `Python documentation`_ for details). Calling this function
+    again after the interpreter has already been initialized is a fatal error.
+
+    If initializing the Python interpreter fails, then the program is terminated.  (This
+    is controlled by the CPython runtime and is an exception to pybind11's normal behavior
+    of throwing exceptions on errors.)
+
+    The remaining optional parameters, `argc`, `argv`, and `add_program_dir_to_path` are
+    used to populate ``sys.argv`` and ``sys.path``.
+    See the |PySys_SetArgvEx documentation|_ for details.
+
+    .. _Python documentation: https://docs.python.org/3/c-api/init.html#c.Py_InitializeEx
+    .. |PySys_SetArgvEx documentation| replace:: ``PySys_SetArgvEx`` documentation
+    .. _PySys_SetArgvEx documentation: https://docs.python.org/3/c-api/init.html#c.PySys_SetArgvEx
+ \endrst */
+inline void initialize_interpreter(bool init_signal_handlers = true,
+                                   int argc = 0,
+                                   const char *const *argv = nullptr,
+                                   bool add_program_dir_to_path = true) {
+#if PY_VERSION_HEX < PYBIND11_PYCONFIG_SUPPORT_PY_VERSION_HEX
+    detail::initialize_interpreter_pre_pyconfig(
+        init_signal_handlers, argc, argv, add_program_dir_to_path);
+#else
+    PyConfig config;
+    PyConfig_InitPythonConfig(&config);
+    // See PR #4473 for background
+    config.parse_argv = 0;
+
+    config.install_signal_handlers = init_signal_handlers ? 1 : 0;
+    initialize_interpreter(&config, argc, argv, add_program_dir_to_path);
+#endif
+}
+
+/** \rst
+    Shut down the Python interpreter. No pybind11 or CPython API functions can be called
+    after this. In addition, pybind11 objects must not outlive the interpreter:
+
+    .. code-block:: cpp
+
+        { // BAD
+            py::initialize_interpreter();
+            auto hello = py::str("Hello, World!");
+            py::finalize_interpreter();
+        } // <-- BOOM, hello's destructor is called after interpreter shutdown
+
+        { // GOOD
+            py::initialize_interpreter();
+            { // scoped
+                auto hello = py::str("Hello, World!");
+            } // <-- OK, hello is cleaned up properly
+            py::finalize_interpreter();
+        }
+
+        { // BETTER
+            py::scoped_interpreter guard{};
+            auto hello = py::str("Hello, World!");
+        }
+
+    .. warning::
+
+        The interpreter can be restarted by calling `initialize_interpreter` again.
+        Modules created using pybind11 can be safely re-initialized. However, Python
+        itself cannot completely unload binary extension modules and there are several
+        caveats with regard to interpreter restarting. All the details can be found
+        in the CPython documentation. In short, not all interpreter memory may be
+        freed, either due to reference cycles or user-created global data.
+
+ \endrst */
+inline void finalize_interpreter() {
+    // Get the internals pointer (without creating it if it doesn't exist).  It's possible for the
+    // internals to be created during Py_Finalize() (e.g. if a py::capsule calls `get_internals()`
+    // during destruction), so we get the pointer-pointer here and check it after Py_Finalize().
+    detail::internals **internals_ptr_ptr = detail::get_internals_pp();
+    // It could also be stashed in state_dict, so look there too:
+    if (object internals_obj
+        = get_internals_obj_from_state_dict(detail::get_python_state_dict())) {
+        internals_ptr_ptr = detail::get_internals_pp_from_capsule(internals_obj);
+    }
+    // Local internals contains data managed by the current interpreter, so we must clear them to
+    // avoid undefined behaviors when initializing another interpreter
+    detail::get_local_internals().registered_types_cpp.clear();
+    detail::get_local_internals().registered_exception_translators.clear();
+
+    Py_Finalize();
+
+    if (internals_ptr_ptr) {
+        delete *internals_ptr_ptr;
+        *internals_ptr_ptr = nullptr;
+    }
+}
+
+/** \rst
+    Scope guard version of `initialize_interpreter` and `finalize_interpreter`.
+    This a move-only guard and only a single instance can exist.
+
+    See `initialize_interpreter` for a discussion of its constructor arguments.
+
+    .. code-block:: cpp
+
+        #include <pybind11/embed.h>
+
+        int main() {
+            py::scoped_interpreter guard{};
+            py::print(Hello, World!);
+        } // <-- interpreter shutdown
+ \endrst */
+class scoped_interpreter {
+public:
+    explicit scoped_interpreter(bool init_signal_handlers = true,
+                                int argc = 0,
+                                const char *const *argv = nullptr,
+                                bool add_program_dir_to_path = true) {
+        initialize_interpreter(init_signal_handlers, argc, argv, add_program_dir_to_path);
+    }
+
+#if PY_VERSION_HEX >= PYBIND11_PYCONFIG_SUPPORT_PY_VERSION_HEX
+    explicit scoped_interpreter(PyConfig *config,
+                                int argc = 0,
+                                const char *const *argv = nullptr,
+                                bool add_program_dir_to_path = true) {
+        initialize_interpreter(config, argc, argv, add_program_dir_to_path);
+    }
+#endif
+
+    scoped_interpreter(const scoped_interpreter &) = delete;
+    scoped_interpreter(scoped_interpreter &&other) noexcept { other.is_valid = false; }
+    scoped_interpreter &operator=(const scoped_interpreter &) = delete;
+    scoped_interpreter &operator=(scoped_interpreter &&) = delete;
+
+    ~scoped_interpreter() {
+        if (is_valid) {
+            finalize_interpreter();
+        }
+    }
+
+private:
+    bool is_valid = true;
+};
+
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/eval.h

@@ -0,0 +1,156 @@
+/*
+    pybind11/eval.h: Support for evaluating Python expressions and statements
+    from strings and files
+
+    Copyright (c) 2016 Klemens Morgenstern <klemens.morgenstern@ed-chemnitz.de> and
+                       Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "pybind11.h"
+
+#include <utility>
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+inline void ensure_builtins_in_globals(object &global) {
+#if defined(PYPY_VERSION) || PY_VERSION_HEX < 0x03080000
+    // Running exec and eval adds `builtins` module under `__builtins__` key to
+    // globals if not yet present.  Python 3.8 made PyRun_String behave
+    // similarly. Let's also do that for older versions, for consistency. This
+    // was missing from PyPy3.8 7.3.7.
+    if (!global.contains("__builtins__"))
+        global["__builtins__"] = module_::import(PYBIND11_BUILTINS_MODULE);
+#else
+    (void) global;
+#endif
+}
+
+PYBIND11_NAMESPACE_END(detail)
+
+enum eval_mode {
+    /// Evaluate a string containing an isolated expression
+    eval_expr,
+
+    /// Evaluate a string containing a single statement. Returns \c none
+    eval_single_statement,
+
+    /// Evaluate a string containing a sequence of statement. Returns \c none
+    eval_statements
+};
+
+template <eval_mode mode = eval_expr>
+object eval(const str &expr, object global = globals(), object local = object()) {
+    if (!local) {
+        local = global;
+    }
+
+    detail::ensure_builtins_in_globals(global);
+
+    /* PyRun_String does not accept a PyObject / encoding specifier,
+       this seems to be the only alternative */
+    std::string buffer = "# -*- coding: utf-8 -*-\n" + (std::string) expr;
+
+    int start = 0;
+    switch (mode) {
+        case eval_expr:
+            start = Py_eval_input;
+            break;
+        case eval_single_statement:
+            start = Py_single_input;
+            break;
+        case eval_statements:
+            start = Py_file_input;
+            break;
+        default:
+            pybind11_fail("invalid evaluation mode");
+    }
+
+    PyObject *result = PyRun_String(buffer.c_str(), start, global.ptr(), local.ptr());
+    if (!result) {
+        throw error_already_set();
+    }
+    return reinterpret_steal<object>(result);
+}
+
+template <eval_mode mode = eval_expr, size_t N>
+object eval(const char (&s)[N], object global = globals(), object local = object()) {
+    /* Support raw string literals by removing common leading whitespace */
+    auto expr = (s[0] == '\n') ? str(module_::import("textwrap").attr("dedent")(s)) : str(s);
+    return eval<mode>(expr, std::move(global), std::move(local));
+}
+
+inline void exec(const str &expr, object global = globals(), object local = object()) {
+    eval<eval_statements>(expr, std::move(global), std::move(local));
+}
+
+template <size_t N>
+void exec(const char (&s)[N], object global = globals(), object local = object()) {
+    eval<eval_statements>(s, std::move(global), std::move(local));
+}
+
+#if defined(PYPY_VERSION)
+template <eval_mode mode = eval_statements>
+object eval_file(str, object, object) {
+    pybind11_fail("eval_file not supported in PyPy3. Use eval");
+}
+template <eval_mode mode = eval_statements>
+object eval_file(str, object) {
+    pybind11_fail("eval_file not supported in PyPy3. Use eval");
+}
+template <eval_mode mode = eval_statements>
+object eval_file(str) {
+    pybind11_fail("eval_file not supported in PyPy3. Use eval");
+}
+#else
+template <eval_mode mode = eval_statements>
+object eval_file(str fname, object global = globals(), object local = object()) {
+    if (!local) {
+        local = global;
+    }
+
+    detail::ensure_builtins_in_globals(global);
+
+    int start = 0;
+    switch (mode) {
+        case eval_expr:
+            start = Py_eval_input;
+            break;
+        case eval_single_statement:
+            start = Py_single_input;
+            break;
+        case eval_statements:
+            start = Py_file_input;
+            break;
+        default:
+            pybind11_fail("invalid evaluation mode");
+    }
+
+    int closeFile = 1;
+    std::string fname_str = (std::string) fname;
+    FILE *f = _Py_fopen_obj(fname.ptr(), "r");
+    if (!f) {
+        PyErr_Clear();
+        pybind11_fail("File \"" + fname_str + "\" could not be opened!");
+    }
+
+    if (!global.contains("__file__")) {
+        global["__file__"] = std::move(fname);
+    }
+
+    PyObject *result
+        = PyRun_FileEx(f, fname_str.c_str(), start, global.ptr(), local.ptr(), closeFile);
+
+    if (!result) {
+        throw error_already_set();
+    }
+    return reinterpret_steal<object>(result);
+}
+#endif
+
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/functional.h

@@ -0,0 +1,149 @@
+/*
+    pybind11/functional.h: std::function<> support
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#define PYBIND11_HAS_TYPE_CASTER_STD_FUNCTION_SPECIALIZATIONS
+
+#include "pybind11.h"
+
+#include <functional>
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+PYBIND11_NAMESPACE_BEGIN(detail)
+PYBIND11_NAMESPACE_BEGIN(type_caster_std_function_specializations)
+
+// ensure GIL is held during functor destruction
+struct func_handle {
+    function f;
+#if !(defined(_MSC_VER) && _MSC_VER == 1916 && defined(PYBIND11_CPP17))
+    // This triggers a syntax error under very special conditions (very weird indeed).
+    explicit
+#endif
+        func_handle(function &&f_) noexcept
+        : f(std::move(f_)) {
+    }
+    func_handle(const func_handle &f_) { operator=(f_); }
+    func_handle &operator=(const func_handle &f_) {
+        gil_scoped_acquire acq;
+        f = f_.f;
+        return *this;
+    }
+    ~func_handle() {
+        gil_scoped_acquire acq;
+        function kill_f(std::move(f));
+    }
+};
+
+// to emulate 'move initialization capture' in C++11
+struct func_wrapper_base {
+    func_handle hfunc;
+    explicit func_wrapper_base(func_handle &&hf) noexcept : hfunc(hf) {}
+};
+
+template <typename Return, typename... Args>
+struct func_wrapper : func_wrapper_base {
+    using func_wrapper_base::func_wrapper_base;
+    Return operator()(Args... args) const {
+        gil_scoped_acquire acq;
+        // casts the returned object as a rvalue to the return type
+        return hfunc.f(std::forward<Args>(args)...).template cast<Return>();
+    }
+};
+
+PYBIND11_NAMESPACE_END(type_caster_std_function_specializations)
+
+template <typename Return, typename... Args>
+struct type_caster<std::function<Return(Args...)>> {
+    using type = std::function<Return(Args...)>;
+    using retval_type = conditional_t<std::is_same<Return, void>::value, void_type, Return>;
+    using function_type = Return (*)(Args...);
+
+public:
+    bool load(handle src, bool convert) {
+        if (src.is_none()) {
+            // Defer accepting None to other overloads (if we aren't in convert mode):
+            if (!convert) {
+                return false;
+            }
+            return true;
+        }
+
+        if (!isinstance<function>(src)) {
+            return false;
+        }
+
+        auto func = reinterpret_borrow<function>(src);
+
+        /*
+           When passing a C++ function as an argument to another C++
+           function via Python, every function call would normally involve
+           a full C++ -> Python -> C++ roundtrip, which can be prohibitive.
+           Here, we try to at least detect the case where the function is
+           stateless (i.e. function pointer or lambda function without
+           captured variables), in which case the roundtrip can be avoided.
+         */
+        if (auto cfunc = func.cpp_function()) {
+            auto *cfunc_self = PyCFunction_GET_SELF(cfunc.ptr());
+            if (cfunc_self == nullptr) {
+                PyErr_Clear();
+            } else if (isinstance<capsule>(cfunc_self)) {
+                auto c = reinterpret_borrow<capsule>(cfunc_self);
+
+                function_record *rec = nullptr;
+                // Check that we can safely reinterpret the capsule into a function_record
+                if (detail::is_function_record_capsule(c)) {
+                    rec = c.get_pointer<function_record>();
+                }
+
+                while (rec != nullptr) {
+                    if (rec->is_stateless
+                        && same_type(typeid(function_type),
+                                     *reinterpret_cast<const std::type_info *>(rec->data[1]))) {
+                        struct capture {
+                            function_type f;
+                        };
+                        value = ((capture *) &rec->data)->f;
+                        return true;
+                    }
+                    rec = rec->next;
+                }
+            }
+            // PYPY segfaults here when passing builtin function like sum.
+            // Raising an fail exception here works to prevent the segfault, but only on gcc.
+            // See PR #1413 for full details
+        }
+
+        value = type_caster_std_function_specializations::func_wrapper<Return, Args...>(
+            type_caster_std_function_specializations::func_handle(std::move(func)));
+        return true;
+    }
+
+    template <typename Func>
+    static handle cast(Func &&f_, return_value_policy policy, handle /* parent */) {
+        if (!f_) {
+            return none().release();
+        }
+
+        auto result = f_.template target<function_type>();
+        if (result) {
+            return cpp_function(*result, policy).release();
+        }
+        return cpp_function(std::forward<Func>(f_), policy).release();
+    }
+
+    PYBIND11_TYPE_CASTER(type,
+                         const_name("Callable[[")
+                             + ::pybind11::detail::concat(make_caster<Args>::name...)
+                             + const_name("], ") + make_caster<retval_type>::name
+                             + const_name("]"));
+};
+
+PYBIND11_NAMESPACE_END(detail)
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/gil.h

@@ -0,0 +1,219 @@
+/*
+    pybind11/gil.h: RAII helpers for managing the GIL
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "detail/common.h"
+
+#include <cassert>
+
+#if !defined(PYBIND11_SIMPLE_GIL_MANAGEMENT)
+#    include "detail/internals.h"
+#endif
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+// forward declarations
+PyThreadState *get_thread_state_unchecked();
+
+PYBIND11_NAMESPACE_END(detail)
+
+#if !defined(PYBIND11_SIMPLE_GIL_MANAGEMENT)
+
+/* The functions below essentially reproduce the PyGILState_* API using a RAII
+ * pattern, but there are a few important differences:
+ *
+ * 1. When acquiring the GIL from an non-main thread during the finalization
+ *    phase, the GILState API blindly terminates the calling thread, which
+ *    is often not what is wanted. This API does not do this.
+ *
+ * 2. The gil_scoped_release function can optionally cut the relationship
+ *    of a PyThreadState and its associated thread, which allows moving it to
+ *    another thread (this is a fairly rare/advanced use case).
+ *
+ * 3. The reference count of an acquired thread state can be controlled. This
+ *    can be handy to prevent cases where callbacks issued from an external
+ *    thread would otherwise constantly construct and destroy thread state data
+ *    structures.
+ *
+ * See the Python bindings of NanoGUI (http://github.com/wjakob/nanogui) for an
+ * example which uses features 2 and 3 to migrate the Python thread of
+ * execution to another thread (to run the event loop on the original thread,
+ * in this case).
+ */
+
+class gil_scoped_acquire {
+public:
+    PYBIND11_NOINLINE gil_scoped_acquire() {
+        auto &internals = detail::get_internals();
+        tstate = (PyThreadState *) PYBIND11_TLS_GET_VALUE(internals.tstate);
+
+        if (!tstate) {
+            /* Check if the GIL was acquired using the PyGILState_* API instead (e.g. if
+               calling from a Python thread). Since we use a different key, this ensures
+               we don't create a new thread state and deadlock in PyEval_AcquireThread
+               below. Note we don't save this state with internals.tstate, since we don't
+               create it we would fail to clear it (its reference count should be > 0). */
+            tstate = PyGILState_GetThisThreadState();
+        }
+
+        if (!tstate) {
+            tstate = PyThreadState_New(internals.istate);
+#    if defined(PYBIND11_DETAILED_ERROR_MESSAGES)
+            if (!tstate) {
+                pybind11_fail("scoped_acquire: could not create thread state!");
+            }
+#    endif
+            tstate->gilstate_counter = 0;
+            PYBIND11_TLS_REPLACE_VALUE(internals.tstate, tstate);
+        } else {
+            release = detail::get_thread_state_unchecked() != tstate;
+        }
+
+        if (release) {
+            PyEval_AcquireThread(tstate);
+        }
+
+        inc_ref();
+    }
+
+    gil_scoped_acquire(const gil_scoped_acquire &) = delete;
+    gil_scoped_acquire &operator=(const gil_scoped_acquire &) = delete;
+
+    void inc_ref() { ++tstate->gilstate_counter; }
+
+    PYBIND11_NOINLINE void dec_ref() {
+        --tstate->gilstate_counter;
+#    if defined(PYBIND11_DETAILED_ERROR_MESSAGES)
+        if (detail::get_thread_state_unchecked() != tstate) {
+            pybind11_fail("scoped_acquire::dec_ref(): thread state must be current!");
+        }
+        if (tstate->gilstate_counter < 0) {
+            pybind11_fail("scoped_acquire::dec_ref(): reference count underflow!");
+        }
+#    endif
+        if (tstate->gilstate_counter == 0) {
+#    if defined(PYBIND11_DETAILED_ERROR_MESSAGES)
+            if (!release) {
+                pybind11_fail("scoped_acquire::dec_ref(): internal error!");
+            }
+#    endif
+            PyThreadState_Clear(tstate);
+            if (active) {
+                PyThreadState_DeleteCurrent();
+            }
+            PYBIND11_TLS_DELETE_VALUE(detail::get_internals().tstate);
+            release = false;
+        }
+    }
+
+    /// This method will disable the PyThreadState_DeleteCurrent call and the
+    /// GIL won't be acquired. This method should be used if the interpreter
+    /// could be shutting down when this is called, as thread deletion is not
+    /// allowed during shutdown. Check _Py_IsFinalizing() on Python 3.7+, and
+    /// protect subsequent code.
+    PYBIND11_NOINLINE void disarm() { active = false; }
+
+    PYBIND11_NOINLINE ~gil_scoped_acquire() {
+        dec_ref();
+        if (release) {
+            PyEval_SaveThread();
+        }
+    }
+
+private:
+    PyThreadState *tstate = nullptr;
+    bool release = true;
+    bool active = true;
+};
+
+class gil_scoped_release {
+public:
+    // PRECONDITION: The GIL must be held when this constructor is called.
+    explicit gil_scoped_release(bool disassoc = false) : disassoc(disassoc) {
+        assert(PyGILState_Check());
+        // `get_internals()` must be called here unconditionally in order to initialize
+        // `internals.tstate` for subsequent `gil_scoped_acquire` calls. Otherwise, an
+        // initialization race could occur as multiple threads try `gil_scoped_acquire`.
+        auto &internals = detail::get_internals();
+        // NOLINTNEXTLINE(cppcoreguidelines-prefer-member-initializer)
+        tstate = PyEval_SaveThread();
+        if (disassoc) {
+            // Python >= 3.7 can remove this, it's an int before 3.7
+            // NOLINTNEXTLINE(readability-qualified-auto)
+            auto key = internals.tstate;
+            PYBIND11_TLS_DELETE_VALUE(key);
+        }
+    }
+
+    gil_scoped_release(const gil_scoped_release &) = delete;
+    gil_scoped_release &operator=(const gil_scoped_release &) = delete;
+
+    /// This method will disable the PyThreadState_DeleteCurrent call and the
+    /// GIL won't be acquired. This method should be used if the interpreter
+    /// could be shutting down when this is called, as thread deletion is not
+    /// allowed during shutdown. Check _Py_IsFinalizing() on Python 3.7+, and
+    /// protect subsequent code.
+    PYBIND11_NOINLINE void disarm() { active = false; }
+
+    ~gil_scoped_release() {
+        if (!tstate) {
+            return;
+        }
+        // `PyEval_RestoreThread()` should not be called if runtime is finalizing
+        if (active) {
+            PyEval_RestoreThread(tstate);
+        }
+        if (disassoc) {
+            // Python >= 3.7 can remove this, it's an int before 3.7
+            // NOLINTNEXTLINE(readability-qualified-auto)
+            auto key = detail::get_internals().tstate;
+            PYBIND11_TLS_REPLACE_VALUE(key, tstate);
+        }
+    }
+
+private:
+    PyThreadState *tstate;
+    bool disassoc;
+    bool active = true;
+};
+
+#else // PYBIND11_SIMPLE_GIL_MANAGEMENT
+
+class gil_scoped_acquire {
+    PyGILState_STATE state;
+
+public:
+    gil_scoped_acquire() : state{PyGILState_Ensure()} {}
+    gil_scoped_acquire(const gil_scoped_acquire &) = delete;
+    gil_scoped_acquire &operator=(const gil_scoped_acquire &) = delete;
+    ~gil_scoped_acquire() { PyGILState_Release(state); }
+    void disarm() {}
+};
+
+class gil_scoped_release {
+    PyThreadState *state;
+
+public:
+    // PRECONDITION: The GIL must be held when this constructor is called.
+    gil_scoped_release() {
+        assert(PyGILState_Check());
+        state = PyEval_SaveThread();
+    }
+    gil_scoped_release(const gil_scoped_release &) = delete;
+    gil_scoped_release &operator=(const gil_scoped_release &) = delete;
+    ~gil_scoped_release() { PyEval_RestoreThread(state); }
+    void disarm() {}
+};
+
+#endif // PYBIND11_SIMPLE_GIL_MANAGEMENT
+
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/gil_safe_call_once.h

@@ -0,0 +1,100 @@
+// Copyright (c) 2023 The pybind Community.
+
+#pragma once
+
+#include "detail/common.h"
+#include "gil.h"
+
+#include <cassert>
+#include <mutex>
+
+#ifdef Py_GIL_DISABLED
+#    include <atomic>
+#endif
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+// Use the `gil_safe_call_once_and_store` class below instead of the naive
+//
+//   static auto imported_obj = py::module_::import("module_name"); // BAD, DO NOT USE!
+//
+// which has two serious issues:
+//
+//     1. Py_DECREF() calls potentially after the Python interpreter was finalized already, and
+//     2. deadlocks in multi-threaded processes (because of missing lock ordering).
+//
+// The following alternative avoids both problems:
+//
+//   PYBIND11_CONSTINIT static py::gil_safe_call_once_and_store<py::object> storage;
+//   auto &imported_obj = storage // Do NOT make this `static`!
+//       .call_once_and_store_result([]() {
+//           return py::module_::import("module_name");
+//       })
+//       .get_stored();
+//
+// The parameter of `call_once_and_store_result()` must be callable. It can make
+// CPython API calls, and in particular, it can temporarily release the GIL.
+//
+// `T` can be any C++ type, it does not have to involve CPython API types.
+//
+// The behavior with regard to signals, e.g. `SIGINT` (`KeyboardInterrupt`),
+// is not ideal. If the main thread is the one to actually run the `Callable`,
+// then a `KeyboardInterrupt` will interrupt it if it is running normal Python
+// code. The situation is different if a non-main thread runs the
+// `Callable`, and then the main thread starts waiting for it to complete:
+// a `KeyboardInterrupt` will not interrupt the non-main thread, but it will
+// get processed only when it is the main thread's turn again and it is running
+// normal Python code. However, this will be unnoticeable for quick call-once
+// functions, which is usually the case.
+template <typename T>
+class gil_safe_call_once_and_store {
+public:
+    // PRECONDITION: The GIL must be held when `call_once_and_store_result()` is called.
+    template <typename Callable>
+    gil_safe_call_once_and_store &call_once_and_store_result(Callable &&fn) {
+        if (!is_initialized_) { // This read is guarded by the GIL.
+            // Multiple threads may enter here, because the GIL is released in the next line and
+            // CPython API calls in the `fn()` call below may release and reacquire the GIL.
+            gil_scoped_release gil_rel; // Needed to establish lock ordering.
+            std::call_once(once_flag_, [&] {
+                // Only one thread will ever enter here.
+                gil_scoped_acquire gil_acq;
+                ::new (storage_) T(fn()); // fn may release, but will reacquire, the GIL.
+                is_initialized_ = true;   // This write is guarded by the GIL.
+            });
+            // All threads will observe `is_initialized_` as true here.
+        }
+        // Intentionally not returning `T &` to ensure the calling code is self-documenting.
+        return *this;
+    }
+
+    // This must only be called after `call_once_and_store_result()` was called.
+    T &get_stored() {
+        assert(is_initialized_);
+        PYBIND11_WARNING_PUSH
+#if !defined(__clang__) && defined(__GNUC__) && __GNUC__ < 5
+        // Needed for gcc 4.8.5
+        PYBIND11_WARNING_DISABLE_GCC("-Wstrict-aliasing")
+#endif
+        return *reinterpret_cast<T *>(storage_);
+        PYBIND11_WARNING_POP
+    }
+
+    constexpr gil_safe_call_once_and_store() = default;
+    PYBIND11_DTOR_CONSTEXPR ~gil_safe_call_once_and_store() = default;
+
+private:
+    alignas(T) char storage_[sizeof(T)] = {};
+    std::once_flag once_flag_ = {};
+#ifdef Py_GIL_DISABLED
+    std::atomic_bool
+#else
+    bool
+#endif
+        is_initialized_{false};
+    // The `is_initialized_`-`storage_` pair is very similar to `std::optional`,
+    // but the latter does not have the triviality properties of former,
+    // therefore `std::optional` is not a viable alternative here.
+};
+
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/iostream.h

@@ -0,0 +1,265 @@
+/*
+    pybind11/iostream.h -- Tools to assist with redirecting cout and cerr to Python
+
+    Copyright (c) 2017 Henry F. Schreiner
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+
+    WARNING: The implementation in this file is NOT thread safe. Multiple
+    threads writing to a redirected ostream concurrently cause data races
+    and potentially buffer overflows. Therefore it is currently a requirement
+    that all (possibly) concurrent redirected ostream writes are protected by
+    a mutex.
+    #HelpAppreciated: Work on iostream.h thread safety.
+    For more background see the discussions under
+    https://github.com/pybind/pybind11/pull/2982 and
+    https://github.com/pybind/pybind11/pull/2995.
+*/
+
+#pragma once
+
+#include "pybind11.h"
+
+#include <algorithm>
+#include <cstring>
+#include <iostream>
+#include <iterator>
+#include <memory>
+#include <ostream>
+#include <streambuf>
+#include <string>
+#include <utility>
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+// Buffer that writes to Python instead of C++
+class pythonbuf : public std::streambuf {
+private:
+    using traits_type = std::streambuf::traits_type;
+
+    const size_t buf_size;
+    std::unique_ptr<char[]> d_buffer;
+    object pywrite;
+    object pyflush;
+
+    int overflow(int c) override {
+        if (!traits_type::eq_int_type(c, traits_type::eof())) {
+            *pptr() = traits_type::to_char_type(c);
+            pbump(1);
+        }
+        return sync() == 0 ? traits_type::not_eof(c) : traits_type::eof();
+    }
+
+    // Computes how many bytes at the end of the buffer are part of an
+    // incomplete sequence of UTF-8 bytes.
+    // Precondition: pbase() < pptr()
+    size_t utf8_remainder() const {
+        const auto rbase = std::reverse_iterator<char *>(pbase());
+        const auto rpptr = std::reverse_iterator<char *>(pptr());
+        auto is_ascii = [](char c) { return (static_cast<unsigned char>(c) & 0x80) == 0x00; };
+        auto is_leading = [](char c) { return (static_cast<unsigned char>(c) & 0xC0) == 0xC0; };
+        auto is_leading_2b = [](char c) { return static_cast<unsigned char>(c) <= 0xDF; };
+        auto is_leading_3b = [](char c) { return static_cast<unsigned char>(c) <= 0xEF; };
+        // If the last character is ASCII, there are no incomplete code points
+        if (is_ascii(*rpptr)) {
+            return 0;
+        }
+        // Otherwise, work back from the end of the buffer and find the first
+        // UTF-8 leading byte
+        const auto rpend = rbase - rpptr >= 3 ? rpptr + 3 : rbase;
+        const auto leading = std::find_if(rpptr, rpend, is_leading);
+        if (leading == rbase) {
+            return 0;
+        }
+        const auto dist = static_cast<size_t>(leading - rpptr);
+        size_t remainder = 0;
+
+        if (dist == 0) {
+            remainder = 1; // 1-byte code point is impossible
+        } else if (dist == 1) {
+            remainder = is_leading_2b(*leading) ? 0 : dist + 1;
+        } else if (dist == 2) {
+            remainder = is_leading_3b(*leading) ? 0 : dist + 1;
+        }
+        // else if (dist >= 3), at least 4 bytes before encountering an UTF-8
+        // leading byte, either no remainder or invalid UTF-8.
+        // Invalid UTF-8 will cause an exception later when converting
+        // to a Python string, so that's not handled here.
+        return remainder;
+    }
+
+    // This function must be non-virtual to be called in a destructor.
+    int _sync() {
+        if (pbase() != pptr()) { // If buffer is not empty
+            gil_scoped_acquire tmp;
+            // This subtraction cannot be negative, so dropping the sign.
+            auto size = static_cast<size_t>(pptr() - pbase());
+            size_t remainder = utf8_remainder();
+
+            if (size > remainder) {
+                str line(pbase(), size - remainder);
+                pywrite(std::move(line));
+                pyflush();
+            }
+
+            // Copy the remainder at the end of the buffer to the beginning:
+            if (remainder > 0) {
+                std::memmove(pbase(), pptr() - remainder, remainder);
+            }
+            setp(pbase(), epptr());
+            pbump(static_cast<int>(remainder));
+        }
+        return 0;
+    }
+
+    int sync() override { return _sync(); }
+
+public:
+    explicit pythonbuf(const object &pyostream, size_t buffer_size = 1024)
+        : buf_size(buffer_size), d_buffer(new char[buf_size]), pywrite(pyostream.attr("write")),
+          pyflush(pyostream.attr("flush")) {
+        setp(d_buffer.get(), d_buffer.get() + buf_size - 1);
+    }
+
+    pythonbuf(pythonbuf &&) = default;
+
+    /// Sync before destroy
+    ~pythonbuf() override { _sync(); }
+};
+
+PYBIND11_NAMESPACE_END(detail)
+
+/** \rst
+    This a move-only guard that redirects output.
+
+    .. code-block:: cpp
+
+        #include <pybind11/iostream.h>
+
+        ...
+
+        {
+            py::scoped_ostream_redirect output;
+            std::cout << "Hello, World!"; // Python stdout
+        } // <-- return std::cout to normal
+
+    You can explicitly pass the c++ stream and the python object,
+    for example to guard stderr instead.
+
+    .. code-block:: cpp
+
+        {
+            py::scoped_ostream_redirect output{
+                std::cerr, py::module::import("sys").attr("stderr")};
+            std::cout << "Hello, World!";
+        }
+ \endrst */
+class scoped_ostream_redirect {
+protected:
+    std::streambuf *old;
+    std::ostream &costream;
+    detail::pythonbuf buffer;
+
+public:
+    explicit scoped_ostream_redirect(std::ostream &costream = std::cout,
+                                     const object &pyostream
+                                     = module_::import("sys").attr("stdout"))
+        : costream(costream), buffer(pyostream) {
+        old = costream.rdbuf(&buffer);
+    }
+
+    ~scoped_ostream_redirect() { costream.rdbuf(old); }
+
+    scoped_ostream_redirect(const scoped_ostream_redirect &) = delete;
+    scoped_ostream_redirect(scoped_ostream_redirect &&other) = default;
+    scoped_ostream_redirect &operator=(const scoped_ostream_redirect &) = delete;
+    scoped_ostream_redirect &operator=(scoped_ostream_redirect &&) = delete;
+};
+
+/** \rst
+    Like `scoped_ostream_redirect`, but redirects cerr by default. This class
+    is provided primary to make ``py::call_guard`` easier to make.
+
+    .. code-block:: cpp
+
+     m.def("noisy_func", &noisy_func,
+           py::call_guard<scoped_ostream_redirect,
+                          scoped_estream_redirect>());
+
+\endrst */
+class scoped_estream_redirect : public scoped_ostream_redirect {
+public:
+    explicit scoped_estream_redirect(std::ostream &costream = std::cerr,
+                                     const object &pyostream
+                                     = module_::import("sys").attr("stderr"))
+        : scoped_ostream_redirect(costream, pyostream) {}
+};
+
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+// Class to redirect output as a context manager. C++ backend.
+class OstreamRedirect {
+    bool do_stdout_;
+    bool do_stderr_;
+    std::unique_ptr<scoped_ostream_redirect> redirect_stdout;
+    std::unique_ptr<scoped_estream_redirect> redirect_stderr;
+
+public:
+    explicit OstreamRedirect(bool do_stdout = true, bool do_stderr = true)
+        : do_stdout_(do_stdout), do_stderr_(do_stderr) {}
+
+    void enter() {
+        if (do_stdout_) {
+            redirect_stdout.reset(new scoped_ostream_redirect());
+        }
+        if (do_stderr_) {
+            redirect_stderr.reset(new scoped_estream_redirect());
+        }
+    }
+
+    void exit() {
+        redirect_stdout.reset();
+        redirect_stderr.reset();
+    }
+};
+
+PYBIND11_NAMESPACE_END(detail)
+
+/** \rst
+    This is a helper function to add a C++ redirect context manager to Python
+    instead of using a C++ guard. To use it, add the following to your binding code:
+
+    .. code-block:: cpp
+
+        #include <pybind11/iostream.h>
+
+        ...
+
+        py::add_ostream_redirect(m, "ostream_redirect");
+
+    You now have a Python context manager that redirects your output:
+
+    .. code-block:: python
+
+        with m.ostream_redirect():
+            m.print_to_cout_function()
+
+    This manager can optionally be told which streams to operate on:
+
+    .. code-block:: python
+
+        with m.ostream_redirect(stdout=true, stderr=true):
+            m.noisy_function_with_error_printing()
+
+ \endrst */
+inline class_<detail::OstreamRedirect>
+add_ostream_redirect(module_ m, const std::string &name = "ostream_redirect") {
+    return class_<detail::OstreamRedirect>(std::move(m), name.c_str(), module_local())
+        .def(init<bool, bool>(), arg("stdout") = true, arg("stderr") = true)
+        .def("__enter__", &detail::OstreamRedirect::enter)
+        .def("__exit__", [](detail::OstreamRedirect &self_, const args &) { self_.exit(); });
+}
+
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/numpy.h

@@ -0,0 +1,2139 @@
+/*
+    pybind11/numpy.h: Basic NumPy support, vectorize() wrapper
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "pybind11.h"
+#include "detail/common.h"
+#include "complex.h"
+#include "gil_safe_call_once.h"
+#include "pytypes.h"
+
+#include <algorithm>
+#include <array>
+#include <cstdint>
+#include <cstdlib>
+#include <cstring>
+#include <functional>
+#include <numeric>
+#include <sstream>
+#include <string>
+#include <type_traits>
+#include <typeindex>
+#include <utility>
+#include <vector>
+
+#if defined(PYBIND11_NUMPY_1_ONLY) && !defined(PYBIND11_INTERNAL_NUMPY_1_ONLY_DETECTED)
+#    error PYBIND11_NUMPY_1_ONLY must be defined before any pybind11 header is included.
+#endif
+
+/* This will be true on all flat address space platforms and allows us to reduce the
+   whole npy_intp / ssize_t / Py_intptr_t business down to just ssize_t for all size
+   and dimension types (e.g. shape, strides, indexing), instead of inflicting this
+   upon the library user.
+   Note that NumPy 2 now uses ssize_t for `npy_intp` to simplify this. */
+static_assert(sizeof(::pybind11::ssize_t) == sizeof(Py_intptr_t), "ssize_t != Py_intptr_t");
+static_assert(std::is_signed<Py_intptr_t>::value, "Py_intptr_t must be signed");
+// We now can reinterpret_cast between py::ssize_t and Py_intptr_t (MSVC + PyPy cares)
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+PYBIND11_WARNING_DISABLE_MSVC(4127)
+
+class dtype; // Forward declaration
+class array; // Forward declaration
+
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+template <>
+struct handle_type_name<dtype> {
+    static constexpr auto name = const_name("numpy.dtype");
+};
+
+template <>
+struct handle_type_name<array> {
+    static constexpr auto name = const_name("numpy.ndarray");
+};
+
+template <typename type, typename SFINAE = void>
+struct npy_format_descriptor;
+
+/* NumPy 1 proxy (always includes legacy fields) */
+struct PyArrayDescr1_Proxy {
+    PyObject_HEAD
+    PyObject *typeobj;
+    char kind;
+    char type;
+    char byteorder;
+    char flags;
+    int type_num;
+    int elsize;
+    int alignment;
+    char *subarray;
+    PyObject *fields;
+    PyObject *names;
+};
+
+#ifndef PYBIND11_NUMPY_1_ONLY
+struct PyArrayDescr_Proxy {
+    PyObject_HEAD
+    PyObject *typeobj;
+    char kind;
+    char type;
+    char byteorder;
+    char _former_flags;
+    int type_num;
+    /* Additional fields are NumPy version specific. */
+};
+#else
+/* NumPy 1.x only, we can expose all fields */
+using PyArrayDescr_Proxy = PyArrayDescr1_Proxy;
+#endif
+
+/* NumPy 2 proxy, including legacy fields */
+struct PyArrayDescr2_Proxy {
+    PyObject_HEAD
+    PyObject *typeobj;
+    char kind;
+    char type;
+    char byteorder;
+    char _former_flags;
+    int type_num;
+    std::uint64_t flags;
+    ssize_t elsize;
+    ssize_t alignment;
+    PyObject *metadata;
+    Py_hash_t hash;
+    void *reserved_null[2];
+    /* The following fields only exist if 0 <= type_num < 2056 */
+    char *subarray;
+    PyObject *fields;
+    PyObject *names;
+};
+
+struct PyArray_Proxy {
+    PyObject_HEAD
+    char *data;
+    int nd;
+    ssize_t *dimensions;
+    ssize_t *strides;
+    PyObject *base;
+    PyObject *descr;
+    int flags;
+};
+
+struct PyVoidScalarObject_Proxy {
+    PyObject_VAR_HEAD char *obval;
+    PyArrayDescr_Proxy *descr;
+    int flags;
+    PyObject *base;
+};
+
+struct numpy_type_info {
+    PyObject *dtype_ptr;
+    std::string format_str;
+};
+
+struct numpy_internals {
+    std::unordered_map<std::type_index, numpy_type_info> registered_dtypes;
+
+    numpy_type_info *get_type_info(const std::type_info &tinfo, bool throw_if_missing = true) {
+        auto it = registered_dtypes.find(std::type_index(tinfo));
+        if (it != registered_dtypes.end()) {
+            return &(it->second);
+        }
+        if (throw_if_missing) {
+            pybind11_fail(std::string("NumPy type info missing for ") + tinfo.name());
+        }
+        return nullptr;
+    }
+
+    template <typename T>
+    numpy_type_info *get_type_info(bool throw_if_missing = true) {
+        return get_type_info(typeid(typename std::remove_cv<T>::type), throw_if_missing);
+    }
+};
+
+PYBIND11_NOINLINE void load_numpy_internals(numpy_internals *&ptr) {
+    ptr = &get_or_create_shared_data<numpy_internals>("_numpy_internals");
+}
+
+inline numpy_internals &get_numpy_internals() {
+    static numpy_internals *ptr = nullptr;
+    if (!ptr) {
+        load_numpy_internals(ptr);
+    }
+    return *ptr;
+}
+
+PYBIND11_NOINLINE module_ import_numpy_core_submodule(const char *submodule_name) {
+    module_ numpy = module_::import("numpy");
+    str version_string = numpy.attr("__version__");
+
+    module_ numpy_lib = module_::import("numpy.lib");
+    object numpy_version = numpy_lib.attr("NumpyVersion")(version_string);
+    int major_version = numpy_version.attr("major").cast<int>();
+
+#ifdef PYBIND11_NUMPY_1_ONLY
+    if (major_version >= 2) {
+        throw std::runtime_error(
+            "This extension was built with PYBIND11_NUMPY_1_ONLY defined, "
+            "but NumPy 2 is used in this process. For NumPy2 compatibility, "
+            "this extension needs to be rebuilt without the PYBIND11_NUMPY_1_ONLY define.");
+    }
+#endif
+    /* `numpy.core` was renamed to `numpy._core` in NumPy 2.0 as it officially
+        became a private module. */
+    std::string numpy_core_path = major_version >= 2 ? "numpy._core" : "numpy.core";
+    return module_::import((numpy_core_path + "." + submodule_name).c_str());
+}
+
+template <typename T>
+struct same_size {
+    template <typename U>
+    using as = bool_constant<sizeof(T) == sizeof(U)>;
+};
+
+template <typename Concrete>
+constexpr int platform_lookup() {
+    return -1;
+}
+
+// Lookup a type according to its size, and return a value corresponding to the NumPy typenum.
+template <typename Concrete, typename T, typename... Ts, typename... Ints>
+constexpr int platform_lookup(int I, Ints... Is) {
+    return sizeof(Concrete) == sizeof(T) ? I : platform_lookup<Concrete, Ts...>(Is...);
+}
+
+struct npy_api {
+    enum constants {
+        NPY_ARRAY_C_CONTIGUOUS_ = 0x0001,
+        NPY_ARRAY_F_CONTIGUOUS_ = 0x0002,
+        NPY_ARRAY_OWNDATA_ = 0x0004,
+        NPY_ARRAY_FORCECAST_ = 0x0010,
+        NPY_ARRAY_ENSUREARRAY_ = 0x0040,
+        NPY_ARRAY_ALIGNED_ = 0x0100,
+        NPY_ARRAY_WRITEABLE_ = 0x0400,
+        NPY_BOOL_ = 0,
+        NPY_BYTE_,
+        NPY_UBYTE_,
+        NPY_SHORT_,
+        NPY_USHORT_,
+        NPY_INT_,
+        NPY_UINT_,
+        NPY_LONG_,
+        NPY_ULONG_,
+        NPY_LONGLONG_,
+        NPY_ULONGLONG_,
+        NPY_FLOAT_,
+        NPY_DOUBLE_,
+        NPY_LONGDOUBLE_,
+        NPY_CFLOAT_,
+        NPY_CDOUBLE_,
+        NPY_CLONGDOUBLE_,
+        NPY_OBJECT_ = 17,
+        NPY_STRING_,
+        NPY_UNICODE_,
+        NPY_VOID_,
+        // Platform-dependent normalization
+        NPY_INT8_ = NPY_BYTE_,
+        NPY_UINT8_ = NPY_UBYTE_,
+        NPY_INT16_ = NPY_SHORT_,
+        NPY_UINT16_ = NPY_USHORT_,
+        // `npy_common.h` defines the integer aliases. In order, it checks:
+        // NPY_BITSOF_LONG, NPY_BITSOF_LONGLONG, NPY_BITSOF_INT, NPY_BITSOF_SHORT, NPY_BITSOF_CHAR
+        // and assigns the alias to the first matching size, so we should check in this order.
+        NPY_INT32_
+        = platform_lookup<std::int32_t, long, int, short>(NPY_LONG_, NPY_INT_, NPY_SHORT_),
+        NPY_UINT32_ = platform_lookup<std::uint32_t, unsigned long, unsigned int, unsigned short>(
+            NPY_ULONG_, NPY_UINT_, NPY_USHORT_),
+        NPY_INT64_
+        = platform_lookup<std::int64_t, long, long long, int>(NPY_LONG_, NPY_LONGLONG_, NPY_INT_),
+        NPY_UINT64_
+        = platform_lookup<std::uint64_t, unsigned long, unsigned long long, unsigned int>(
+            NPY_ULONG_, NPY_ULONGLONG_, NPY_UINT_),
+    };
+
+    unsigned int PyArray_RUNTIME_VERSION_;
+
+    struct PyArray_Dims {
+        Py_intptr_t *ptr;
+        int len;
+    };
+
+    static npy_api &get() {
+        PYBIND11_CONSTINIT static gil_safe_call_once_and_store<npy_api> storage;
+        return storage.call_once_and_store_result(lookup).get_stored();
+    }
+
+    bool PyArray_Check_(PyObject *obj) const {
+        return PyObject_TypeCheck(obj, PyArray_Type_) != 0;
+    }
+    bool PyArrayDescr_Check_(PyObject *obj) const {
+        return PyObject_TypeCheck(obj, PyArrayDescr_Type_) != 0;
+    }
+
+    unsigned int (*PyArray_GetNDArrayCFeatureVersion_)();
+    PyObject *(*PyArray_DescrFromType_)(int);
+    PyObject *(*PyArray_NewFromDescr_)(PyTypeObject *,
+                                       PyObject *,
+                                       int,
+                                       Py_intptr_t const *,
+                                       Py_intptr_t const *,
+                                       void *,
+                                       int,
+                                       PyObject *);
+    // Unused. Not removed because that affects ABI of the class.
+    PyObject *(*PyArray_DescrNewFromType_)(int);
+    int (*PyArray_CopyInto_)(PyObject *, PyObject *);
+    PyObject *(*PyArray_NewCopy_)(PyObject *, int);
+    PyTypeObject *PyArray_Type_;
+    PyTypeObject *PyVoidArrType_Type_;
+    PyTypeObject *PyArrayDescr_Type_;
+    PyObject *(*PyArray_DescrFromScalar_)(PyObject *);
+    PyObject *(*PyArray_FromAny_)(PyObject *, PyObject *, int, int, int, PyObject *);
+    int (*PyArray_DescrConverter_)(PyObject *, PyObject **);
+    bool (*PyArray_EquivTypes_)(PyObject *, PyObject *);
+#ifdef PYBIND11_NUMPY_1_ONLY
+    int (*PyArray_GetArrayParamsFromObject_)(PyObject *,
+                                             PyObject *,
+                                             unsigned char,
+                                             PyObject **,
+                                             int *,
+                                             Py_intptr_t *,
+                                             PyObject **,
+                                             PyObject *);
+#endif
+    PyObject *(*PyArray_Squeeze_)(PyObject *);
+    // Unused. Not removed because that affects ABI of the class.
+    int (*PyArray_SetBaseObject_)(PyObject *, PyObject *);
+    PyObject *(*PyArray_Resize_)(PyObject *, PyArray_Dims *, int, int);
+    PyObject *(*PyArray_Newshape_)(PyObject *, PyArray_Dims *, int);
+    PyObject *(*PyArray_View_)(PyObject *, PyObject *, PyObject *);
+
+private:
+    enum functions {
+        API_PyArray_GetNDArrayCFeatureVersion = 211,
+        API_PyArray_Type = 2,
+        API_PyArrayDescr_Type = 3,
+        API_PyVoidArrType_Type = 39,
+        API_PyArray_DescrFromType = 45,
+        API_PyArray_DescrFromScalar = 57,
+        API_PyArray_FromAny = 69,
+        API_PyArray_Resize = 80,
+        // CopyInto was slot 82 and 50 was effectively an alias. NumPy 2 removed 82.
+        API_PyArray_CopyInto = 50,
+        API_PyArray_NewCopy = 85,
+        API_PyArray_NewFromDescr = 94,
+        API_PyArray_DescrNewFromType = 96,
+        API_PyArray_Newshape = 135,
+        API_PyArray_Squeeze = 136,
+        API_PyArray_View = 137,
+        API_PyArray_DescrConverter = 174,
+        API_PyArray_EquivTypes = 182,
+#ifdef PYBIND11_NUMPY_1_ONLY
+        API_PyArray_GetArrayParamsFromObject = 278,
+#endif
+        API_PyArray_SetBaseObject = 282
+    };
+
+    static npy_api lookup() {
+        module_ m = detail::import_numpy_core_submodule("multiarray");
+        auto c = m.attr("_ARRAY_API");
+        void **api_ptr = (void **) PyCapsule_GetPointer(c.ptr(), nullptr);
+        if (api_ptr == nullptr) {
+            raise_from(PyExc_SystemError, "FAILURE obtaining numpy _ARRAY_API pointer.");
+            throw error_already_set();
+        }
+        npy_api api;
+#define DECL_NPY_API(Func) api.Func##_ = (decltype(api.Func##_)) api_ptr[API_##Func];
+        DECL_NPY_API(PyArray_GetNDArrayCFeatureVersion);
+        api.PyArray_RUNTIME_VERSION_ = api.PyArray_GetNDArrayCFeatureVersion_();
+        if (api.PyArray_RUNTIME_VERSION_ < 0x7) {
+            pybind11_fail("pybind11 numpy support requires numpy >= 1.7.0");
+        }
+        DECL_NPY_API(PyArray_Type);
+        DECL_NPY_API(PyVoidArrType_Type);
+        DECL_NPY_API(PyArrayDescr_Type);
+        DECL_NPY_API(PyArray_DescrFromType);
+        DECL_NPY_API(PyArray_DescrFromScalar);
+        DECL_NPY_API(PyArray_FromAny);
+        DECL_NPY_API(PyArray_Resize);
+        DECL_NPY_API(PyArray_CopyInto);
+        DECL_NPY_API(PyArray_NewCopy);
+        DECL_NPY_API(PyArray_NewFromDescr);
+        DECL_NPY_API(PyArray_DescrNewFromType);
+        DECL_NPY_API(PyArray_Newshape);
+        DECL_NPY_API(PyArray_Squeeze);
+        DECL_NPY_API(PyArray_View);
+        DECL_NPY_API(PyArray_DescrConverter);
+        DECL_NPY_API(PyArray_EquivTypes);
+#ifdef PYBIND11_NUMPY_1_ONLY
+        DECL_NPY_API(PyArray_GetArrayParamsFromObject);
+#endif
+        DECL_NPY_API(PyArray_SetBaseObject);
+
+#undef DECL_NPY_API
+        return api;
+    }
+};
+
+inline PyArray_Proxy *array_proxy(void *ptr) { return reinterpret_cast<PyArray_Proxy *>(ptr); }
+
+inline const PyArray_Proxy *array_proxy(const void *ptr) {
+    return reinterpret_cast<const PyArray_Proxy *>(ptr);
+}
+
+inline PyArrayDescr_Proxy *array_descriptor_proxy(PyObject *ptr) {
+    return reinterpret_cast<PyArrayDescr_Proxy *>(ptr);
+}
+
+inline const PyArrayDescr_Proxy *array_descriptor_proxy(const PyObject *ptr) {
+    return reinterpret_cast<const PyArrayDescr_Proxy *>(ptr);
+}
+
+inline const PyArrayDescr1_Proxy *array_descriptor1_proxy(const PyObject *ptr) {
+    return reinterpret_cast<const PyArrayDescr1_Proxy *>(ptr);
+}
+
+inline const PyArrayDescr2_Proxy *array_descriptor2_proxy(const PyObject *ptr) {
+    return reinterpret_cast<const PyArrayDescr2_Proxy *>(ptr);
+}
+
+inline bool check_flags(const void *ptr, int flag) {
+    return (flag == (array_proxy(ptr)->flags & flag));
+}
+
+template <typename T>
+struct is_std_array : std::false_type {};
+template <typename T, size_t N>
+struct is_std_array<std::array<T, N>> : std::true_type {};
+template <typename T>
+struct is_complex : std::false_type {};
+template <typename T>
+struct is_complex<std::complex<T>> : std::true_type {};
+
+template <typename T>
+struct array_info_scalar {
+    using type = T;
+    static constexpr bool is_array = false;
+    static constexpr bool is_empty = false;
+    static constexpr auto extents = const_name("");
+    static void append_extents(list & /* shape */) {}
+};
+// Computes underlying type and a comma-separated list of extents for array
+// types (any mix of std::array and built-in arrays). An array of char is
+// treated as scalar because it gets special handling.
+template <typename T>
+struct array_info : array_info_scalar<T> {};
+template <typename T, size_t N>
+struct array_info<std::array<T, N>> {
+    using type = typename array_info<T>::type;
+    static constexpr bool is_array = true;
+    static constexpr bool is_empty = (N == 0) || array_info<T>::is_empty;
+    static constexpr size_t extent = N;
+
+    // appends the extents to shape
+    static void append_extents(list &shape) {
+        shape.append(N);
+        array_info<T>::append_extents(shape);
+    }
+
+    static constexpr auto extents = const_name<array_info<T>::is_array>(
+        ::pybind11::detail::concat(const_name<N>(), array_info<T>::extents), const_name<N>());
+};
+// For numpy we have special handling for arrays of characters, so we don't include
+// the size in the array extents.
+template <size_t N>
+struct array_info<char[N]> : array_info_scalar<char[N]> {};
+template <size_t N>
+struct array_info<std::array<char, N>> : array_info_scalar<std::array<char, N>> {};
+template <typename T, size_t N>
+struct array_info<T[N]> : array_info<std::array<T, N>> {};
+template <typename T>
+using remove_all_extents_t = typename array_info<T>::type;
+
+template <typename T>
+using is_pod_struct
+    = all_of<std::is_standard_layout<T>, // since we're accessing directly in memory
+                                         // we need a standard layout type
+#if defined(__GLIBCXX__)                                                                          \
+    && (__GLIBCXX__ < 20150422 || __GLIBCXX__ == 20150426 || __GLIBCXX__ == 20150623              \
+        || __GLIBCXX__ == 20150626 || __GLIBCXX__ == 20160803)
+             // libstdc++ < 5 (including versions 4.8.5, 4.9.3 and 4.9.4 which were released after
+             // 5) don't implement is_trivially_copyable, so approximate it
+             std::is_trivially_destructible<T>,
+             satisfies_any_of<T, std::has_trivial_copy_constructor, std::has_trivial_copy_assign>,
+#else
+             std::is_trivially_copyable<T>,
+#endif
+             satisfies_none_of<T,
+                               std::is_reference,
+                               std::is_array,
+                               is_std_array,
+                               std::is_arithmetic,
+                               is_complex,
+                               std::is_enum>>;
+
+// Replacement for std::is_pod (deprecated in C++20)
+template <typename T>
+using is_pod = all_of<std::is_standard_layout<T>, std::is_trivial<T>>;
+
+template <ssize_t Dim = 0, typename Strides>
+ssize_t byte_offset_unsafe(const Strides &) {
+    return 0;
+}
+template <ssize_t Dim = 0, typename Strides, typename... Ix>
+ssize_t byte_offset_unsafe(const Strides &strides, ssize_t i, Ix... index) {
+    return i * strides[Dim] + byte_offset_unsafe<Dim + 1>(strides, index...);
+}
+
+/**
+ * Proxy class providing unsafe, unchecked const access to array data.  This is constructed through
+ * the `unchecked<T, N>()` method of `array` or the `unchecked<N>()` method of `array_t<T>`. `Dims`
+ * will be -1 for dimensions determined at runtime.
+ */
+template <typename T, ssize_t Dims>
+class unchecked_reference {
+protected:
+    static constexpr bool Dynamic = Dims < 0;
+    const unsigned char *data_;
+    // Storing the shape & strides in local variables (i.e. these arrays) allows the compiler to
+    // make large performance gains on big, nested loops, but requires compile-time dimensions
+    conditional_t<Dynamic, const ssize_t *, std::array<ssize_t, (size_t) Dims>> shape_, strides_;
+    const ssize_t dims_;
+
+    friend class pybind11::array;
+    // Constructor for compile-time dimensions:
+    template <bool Dyn = Dynamic>
+    unchecked_reference(const void *data,
+                        const ssize_t *shape,
+                        const ssize_t *strides,
+                        enable_if_t<!Dyn, ssize_t>)
+        : data_{reinterpret_cast<const unsigned char *>(data)}, dims_{Dims} {
+        for (size_t i = 0; i < (size_t) dims_; i++) {
+            shape_[i] = shape[i];
+            strides_[i] = strides[i];
+        }
+    }
+    // Constructor for runtime dimensions:
+    template <bool Dyn = Dynamic>
+    unchecked_reference(const void *data,
+                        const ssize_t *shape,
+                        const ssize_t *strides,
+                        enable_if_t<Dyn, ssize_t> dims)
+        : data_{reinterpret_cast<const unsigned char *>(data)}, shape_{shape}, strides_{strides},
+          dims_{dims} {}
+
+public:
+    /**
+     * Unchecked const reference access to data at the given indices.  For a compile-time known
+     * number of dimensions, this requires the correct number of arguments; for run-time
+     * dimensionality, this is not checked (and so is up to the caller to use safely).
+     */
+    template <typename... Ix>
+    const T &operator()(Ix... index) const {
+        static_assert(ssize_t{sizeof...(Ix)} == Dims || Dynamic,
+                      "Invalid number of indices for unchecked array reference");
+        return *reinterpret_cast<const T *>(data_
+                                            + byte_offset_unsafe(strides_, ssize_t(index)...));
+    }
+    /**
+     * Unchecked const reference access to data; this operator only participates if the reference
+     * is to a 1-dimensional array.  When present, this is exactly equivalent to `obj(index)`.
+     */
+    template <ssize_t D = Dims, typename = enable_if_t<D == 1 || Dynamic>>
+    const T &operator[](ssize_t index) const {
+        return operator()(index);
+    }
+
+    /// Pointer access to the data at the given indices.
+    template <typename... Ix>
+    const T *data(Ix... ix) const {
+        return &operator()(ssize_t(ix)...);
+    }
+
+    /// Returns the item size, i.e. sizeof(T)
+    constexpr static ssize_t itemsize() { return sizeof(T); }
+
+    /// Returns the shape (i.e. size) of dimension `dim`
+    ssize_t shape(ssize_t dim) const { return shape_[(size_t) dim]; }
+
+    /// Returns the number of dimensions of the array
+    ssize_t ndim() const { return dims_; }
+
+    /// Returns the total number of elements in the referenced array, i.e. the product of the
+    /// shapes
+    template <bool Dyn = Dynamic>
+    enable_if_t<!Dyn, ssize_t> size() const {
+        return std::accumulate(
+            shape_.begin(), shape_.end(), (ssize_t) 1, std::multiplies<ssize_t>());
+    }
+    template <bool Dyn = Dynamic>
+    enable_if_t<Dyn, ssize_t> size() const {
+        return std::accumulate(shape_, shape_ + ndim(), (ssize_t) 1, std::multiplies<ssize_t>());
+    }
+
+    /// Returns the total number of bytes used by the referenced data.  Note that the actual span
+    /// in memory may be larger if the referenced array has non-contiguous strides (e.g. for a
+    /// slice).
+    ssize_t nbytes() const { return size() * itemsize(); }
+};
+
+template <typename T, ssize_t Dims>
+class unchecked_mutable_reference : public unchecked_reference<T, Dims> {
+    friend class pybind11::array;
+    using ConstBase = unchecked_reference<T, Dims>;
+    using ConstBase::ConstBase;
+    using ConstBase::Dynamic;
+
+public:
+    // Bring in const-qualified versions from base class
+    using ConstBase::operator();
+    using ConstBase::operator[];
+
+    /// Mutable, unchecked access to data at the given indices.
+    template <typename... Ix>
+    T &operator()(Ix... index) {
+        static_assert(ssize_t{sizeof...(Ix)} == Dims || Dynamic,
+                      "Invalid number of indices for unchecked array reference");
+        return const_cast<T &>(ConstBase::operator()(index...));
+    }
+    /**
+     * Mutable, unchecked access data at the given index; this operator only participates if the
+     * reference is to a 1-dimensional array (or has runtime dimensions).  When present, this is
+     * exactly equivalent to `obj(index)`.
+     */
+    template <ssize_t D = Dims, typename = enable_if_t<D == 1 || Dynamic>>
+    T &operator[](ssize_t index) {
+        return operator()(index);
+    }
+
+    /// Mutable pointer access to the data at the given indices.
+    template <typename... Ix>
+    T *mutable_data(Ix... ix) {
+        return &operator()(ssize_t(ix)...);
+    }
+};
+
+template <typename T, ssize_t Dim>
+struct type_caster<unchecked_reference<T, Dim>> {
+    static_assert(Dim == 0 && Dim > 0 /* always fail */,
+                  "unchecked array proxy object is not castable");
+};
+template <typename T, ssize_t Dim>
+struct type_caster<unchecked_mutable_reference<T, Dim>>
+    : type_caster<unchecked_reference<T, Dim>> {};
+
+PYBIND11_NAMESPACE_END(detail)
+
+class dtype : public object {
+public:
+    PYBIND11_OBJECT_DEFAULT(dtype, object, detail::npy_api::get().PyArrayDescr_Check_)
+
+    explicit dtype(const buffer_info &info) {
+        dtype descr(_dtype_from_pep3118()(pybind11::str(info.format)));
+        // If info.itemsize == 0, use the value calculated from the format string
+        m_ptr = descr.strip_padding(info.itemsize != 0 ? info.itemsize : descr.itemsize())
+                    .release()
+                    .ptr();
+    }
+
+    explicit dtype(const pybind11::str &format) : dtype(from_args(format)) {}
+
+    explicit dtype(const std::string &format) : dtype(pybind11::str(format)) {}
+
+    explicit dtype(const char *format) : dtype(pybind11::str(format)) {}
+
+    dtype(list names, list formats, list offsets, ssize_t itemsize) {
+        dict args;
+        args["names"] = std::move(names);
+        args["formats"] = std::move(formats);
+        args["offsets"] = std::move(offsets);
+        args["itemsize"] = pybind11::int_(itemsize);
+        m_ptr = from_args(args).release().ptr();
+    }
+
+    /// Return dtype for the given typenum (one of the NPY_TYPES).
+    /// https://numpy.org/devdocs/reference/c-api/array.html#c.PyArray_DescrFromType
+    explicit dtype(int typenum)
+        : object(detail::npy_api::get().PyArray_DescrFromType_(typenum), stolen_t{}) {
+        if (m_ptr == nullptr) {
+            throw error_already_set();
+        }
+    }
+
+    /// This is essentially the same as calling numpy.dtype(args) in Python.
+    static dtype from_args(const object &args) {
+        PyObject *ptr = nullptr;
+        if ((detail::npy_api::get().PyArray_DescrConverter_(args.ptr(), &ptr) == 0) || !ptr) {
+            throw error_already_set();
+        }
+        return reinterpret_steal<dtype>(ptr);
+    }
+
+    /// Return dtype associated with a C++ type.
+    template <typename T>
+    static dtype of() {
+        return detail::npy_format_descriptor<typename std::remove_cv<T>::type>::dtype();
+    }
+
+    /// Size of the data type in bytes.
+#ifdef PYBIND11_NUMPY_1_ONLY
+    ssize_t itemsize() const { return detail::array_descriptor_proxy(m_ptr)->elsize; }
+#else
+    ssize_t itemsize() const {
+        if (detail::npy_api::get().PyArray_RUNTIME_VERSION_ < 0x12) {
+            return detail::array_descriptor1_proxy(m_ptr)->elsize;
+        }
+        return detail::array_descriptor2_proxy(m_ptr)->elsize;
+    }
+#endif
+
+    /// Returns true for structured data types.
+#ifdef PYBIND11_NUMPY_1_ONLY
+    bool has_fields() const { return detail::array_descriptor_proxy(m_ptr)->names != nullptr; }
+#else
+    bool has_fields() const {
+        if (detail::npy_api::get().PyArray_RUNTIME_VERSION_ < 0x12) {
+            return detail::array_descriptor1_proxy(m_ptr)->names != nullptr;
+        }
+        const auto *proxy = detail::array_descriptor2_proxy(m_ptr);
+        if (proxy->type_num < 0 || proxy->type_num >= 2056) {
+            return false;
+        }
+        return proxy->names != nullptr;
+    }
+#endif
+
+    /// Single-character code for dtype's kind.
+    /// For example, floating point types are 'f' and integral types are 'i'.
+    char kind() const { return detail::array_descriptor_proxy(m_ptr)->kind; }
+
+    /// Single-character for dtype's type.
+    /// For example, ``float`` is 'f', ``double`` 'd', ``int`` 'i', and ``long`` 'l'.
+    char char_() const {
+        // Note: The signature, `dtype::char_` follows the naming of NumPy's
+        // public Python API (i.e., ``dtype.char``), rather than its internal
+        // C API (``PyArray_Descr::type``).
+        return detail::array_descriptor_proxy(m_ptr)->type;
+    }
+
+    /// type number of dtype.
+    int num() const {
+        // Note: The signature, `dtype::num` follows the naming of NumPy's public
+        // Python API (i.e., ``dtype.num``), rather than its internal
+        // C API (``PyArray_Descr::type_num``).
+        return detail::array_descriptor_proxy(m_ptr)->type_num;
+    }
+
+    /// Single character for byteorder
+    char byteorder() const { return detail::array_descriptor_proxy(m_ptr)->byteorder; }
+
+/// Alignment of the data type
+#ifdef PYBIND11_NUMPY_1_ONLY
+    int alignment() const { return detail::array_descriptor_proxy(m_ptr)->alignment; }
+#else
+    ssize_t alignment() const {
+        if (detail::npy_api::get().PyArray_RUNTIME_VERSION_ < 0x12) {
+            return detail::array_descriptor1_proxy(m_ptr)->alignment;
+        }
+        return detail::array_descriptor2_proxy(m_ptr)->alignment;
+    }
+#endif
+
+/// Flags for the array descriptor
+#ifdef PYBIND11_NUMPY_1_ONLY
+    char flags() const { return detail::array_descriptor_proxy(m_ptr)->flags; }
+#else
+    std::uint64_t flags() const {
+        if (detail::npy_api::get().PyArray_RUNTIME_VERSION_ < 0x12) {
+            return (unsigned char) detail::array_descriptor1_proxy(m_ptr)->flags;
+        }
+        return detail::array_descriptor2_proxy(m_ptr)->flags;
+    }
+#endif
+
+private:
+    static object &_dtype_from_pep3118() {
+        PYBIND11_CONSTINIT static gil_safe_call_once_and_store<object> storage;
+        return storage
+            .call_once_and_store_result([]() {
+                return detail::import_numpy_core_submodule("_internal")
+                    .attr("_dtype_from_pep3118");
+            })
+            .get_stored();
+    }
+
+    dtype strip_padding(ssize_t itemsize) {
+        // Recursively strip all void fields with empty names that are generated for
+        // padding fields (as of NumPy v1.11).
+        if (!has_fields()) {
+            return *this;
+        }
+
+        struct field_descr {
+            pybind11::str name;
+            object format;
+            pybind11::int_ offset;
+            field_descr(pybind11::str &&name, object &&format, pybind11::int_ &&offset)
+                : name{std::move(name)}, format{std::move(format)}, offset{std::move(offset)} {};
+        };
+        auto field_dict = attr("fields").cast<dict>();
+        std::vector<field_descr> field_descriptors;
+        field_descriptors.reserve(field_dict.size());
+
+        for (auto field : field_dict.attr("items")()) {
+            auto spec = field.cast<tuple>();
+            auto name = spec[0].cast<pybind11::str>();
+            auto spec_fo = spec[1].cast<tuple>();
+            auto format = spec_fo[0].cast<dtype>();
+            auto offset = spec_fo[1].cast<pybind11::int_>();
+            if ((len(name) == 0u) && format.kind() == 'V') {
+                continue;
+            }
+            field_descriptors.emplace_back(
+                std::move(name), format.strip_padding(format.itemsize()), std::move(offset));
+        }
+
+        std::sort(field_descriptors.begin(),
+                  field_descriptors.end(),
+                  [](const field_descr &a, const field_descr &b) {
+                      return a.offset.cast<int>() < b.offset.cast<int>();
+                  });
+
+        list names, formats, offsets;
+        for (auto &descr : field_descriptors) {
+            names.append(std::move(descr.name));
+            formats.append(std::move(descr.format));
+            offsets.append(std::move(descr.offset));
+        }
+        return dtype(std::move(names), std::move(formats), std::move(offsets), itemsize);
+    }
+};
+
+class array : public buffer {
+public:
+    PYBIND11_OBJECT_CVT(array, buffer, detail::npy_api::get().PyArray_Check_, raw_array)
+
+    enum {
+        c_style = detail::npy_api::NPY_ARRAY_C_CONTIGUOUS_,
+        f_style = detail::npy_api::NPY_ARRAY_F_CONTIGUOUS_,
+        forcecast = detail::npy_api::NPY_ARRAY_FORCECAST_
+    };
+
+    array() : array(0, static_cast<const double *>(nullptr)) {}
+
+    using ShapeContainer = detail::any_container<ssize_t>;
+    using StridesContainer = detail::any_container<ssize_t>;
+
+    // Constructs an array taking shape/strides from arbitrary container types
+    array(const pybind11::dtype &dt,
+          ShapeContainer shape,
+          StridesContainer strides,
+          const void *ptr = nullptr,
+          handle base = handle()) {
+
+        if (strides->empty()) {
+            *strides = detail::c_strides(*shape, dt.itemsize());
+        }
+
+        auto ndim = shape->size();
+        if (ndim != strides->size()) {
+            pybind11_fail("NumPy: shape ndim doesn't match strides ndim");
+        }
+        auto descr = dt;
+
+        int flags = 0;
+        if (base && ptr) {
+            if (isinstance<array>(base)) {
+                /* Copy flags from base (except ownership bit) */
+                flags = reinterpret_borrow<array>(base).flags()
+                        & ~detail::npy_api::NPY_ARRAY_OWNDATA_;
+            } else {
+                /* Writable by default, easy to downgrade later on if needed */
+                flags = detail::npy_api::NPY_ARRAY_WRITEABLE_;
+            }
+        }
+
+        auto &api = detail::npy_api::get();
+        auto tmp = reinterpret_steal<object>(api.PyArray_NewFromDescr_(
+            api.PyArray_Type_,
+            descr.release().ptr(),
+            (int) ndim,
+            // Use reinterpret_cast for PyPy on Windows (remove if fixed, checked on 7.3.1)
+            reinterpret_cast<Py_intptr_t *>(shape->data()),
+            reinterpret_cast<Py_intptr_t *>(strides->data()),
+            const_cast<void *>(ptr),
+            flags,
+            nullptr));
+        if (!tmp) {
+            throw error_already_set();
+        }
+        if (ptr) {
+            if (base) {
+                api.PyArray_SetBaseObject_(tmp.ptr(), base.inc_ref().ptr());
+            } else {
+                tmp = reinterpret_steal<object>(
+                    api.PyArray_NewCopy_(tmp.ptr(), -1 /* any order */));
+            }
+        }
+        m_ptr = tmp.release().ptr();
+    }
+
+    array(const pybind11::dtype &dt,
+          ShapeContainer shape,
+          const void *ptr = nullptr,
+          handle base = handle())
+        : array(dt, std::move(shape), {}, ptr, base) {}
+
+    template <typename T,
+              typename
+              = detail::enable_if_t<std::is_integral<T>::value && !std::is_same<bool, T>::value>>
+    array(const pybind11::dtype &dt, T count, const void *ptr = nullptr, handle base = handle())
+        : array(dt, {{count}}, ptr, base) {}
+
+    template <typename T>
+    array(ShapeContainer shape, StridesContainer strides, const T *ptr, handle base = handle())
+        : array(pybind11::dtype::of<T>(),
+                std::move(shape),
+                std::move(strides),
+                reinterpret_cast<const void *>(ptr),
+                base) {}
+
+    template <typename T>
+    array(ShapeContainer shape, const T *ptr, handle base = handle())
+        : array(std::move(shape), {}, ptr, base) {}
+
+    template <typename T>
+    explicit array(ssize_t count, const T *ptr, handle base = handle())
+        : array({count}, {}, ptr, base) {}
+
+    explicit array(const buffer_info &info, handle base = handle())
+        : array(pybind11::dtype(info), info.shape, info.strides, info.ptr, base) {}
+
+    /// Array descriptor (dtype)
+    pybind11::dtype dtype() const {
+        return reinterpret_borrow<pybind11::dtype>(detail::array_proxy(m_ptr)->descr);
+    }
+
+    /// Total number of elements
+    ssize_t size() const {
+        return std::accumulate(shape(), shape() + ndim(), (ssize_t) 1, std::multiplies<ssize_t>());
+    }
+
+    /// Byte size of a single element
+    ssize_t itemsize() const { return dtype().itemsize(); }
+
+    /// Total number of bytes
+    ssize_t nbytes() const { return size() * itemsize(); }
+
+    /// Number of dimensions
+    ssize_t ndim() const { return detail::array_proxy(m_ptr)->nd; }
+
+    /// Base object
+    object base() const { return reinterpret_borrow<object>(detail::array_proxy(m_ptr)->base); }
+
+    /// Dimensions of the array
+    const ssize_t *shape() const { return detail::array_proxy(m_ptr)->dimensions; }
+
+    /// Dimension along a given axis
+    ssize_t shape(ssize_t dim) const {
+        if (dim >= ndim()) {
+            fail_dim_check(dim, "invalid axis");
+        }
+        return shape()[dim];
+    }
+
+    /// Strides of the array
+    const ssize_t *strides() const { return detail::array_proxy(m_ptr)->strides; }
+
+    /// Stride along a given axis
+    ssize_t strides(ssize_t dim) const {
+        if (dim >= ndim()) {
+            fail_dim_check(dim, "invalid axis");
+        }
+        return strides()[dim];
+    }
+
+    /// Return the NumPy array flags
+    int flags() const { return detail::array_proxy(m_ptr)->flags; }
+
+    /// If set, the array is writeable (otherwise the buffer is read-only)
+    bool writeable() const {
+        return detail::check_flags(m_ptr, detail::npy_api::NPY_ARRAY_WRITEABLE_);
+    }
+
+    /// If set, the array owns the data (will be freed when the array is deleted)
+    bool owndata() const {
+        return detail::check_flags(m_ptr, detail::npy_api::NPY_ARRAY_OWNDATA_);
+    }
+
+    /// Pointer to the contained data. If index is not provided, points to the
+    /// beginning of the buffer. May throw if the index would lead to out of bounds access.
+    template <typename... Ix>
+    const void *data(Ix... index) const {
+        return static_cast<const void *>(detail::array_proxy(m_ptr)->data + offset_at(index...));
+    }
+
+    /// Mutable pointer to the contained data. If index is not provided, points to the
+    /// beginning of the buffer. May throw if the index would lead to out of bounds access.
+    /// May throw if the array is not writeable.
+    template <typename... Ix>
+    void *mutable_data(Ix... index) {
+        check_writeable();
+        return static_cast<void *>(detail::array_proxy(m_ptr)->data + offset_at(index...));
+    }
+
+    /// Byte offset from beginning of the array to a given index (full or partial).
+    /// May throw if the index would lead to out of bounds access.
+    template <typename... Ix>
+    ssize_t offset_at(Ix... index) const {
+        if ((ssize_t) sizeof...(index) > ndim()) {
+            fail_dim_check(sizeof...(index), "too many indices for an array");
+        }
+        return byte_offset(ssize_t(index)...);
+    }
+
+    ssize_t offset_at() const { return 0; }
+
+    /// Item count from beginning of the array to a given index (full or partial).
+    /// May throw if the index would lead to out of bounds access.
+    template <typename... Ix>
+    ssize_t index_at(Ix... index) const {
+        return offset_at(index...) / itemsize();
+    }
+
+    /**
+     * Returns a proxy object that provides access to the array's data without bounds or
+     * dimensionality checking.  Will throw if the array is missing the `writeable` flag.  Use with
+     * care: the array must not be destroyed or reshaped for the duration of the returned object,
+     * and the caller must take care not to access invalid dimensions or dimension indices.
+     */
+    template <typename T, ssize_t Dims = -1>
+    detail::unchecked_mutable_reference<T, Dims> mutable_unchecked() & {
+        if (Dims >= 0 && ndim() != Dims) {
+            throw std::domain_error("array has incorrect number of dimensions: "
+                                    + std::to_string(ndim()) + "; expected "
+                                    + std::to_string(Dims));
+        }
+        return detail::unchecked_mutable_reference<T, Dims>(
+            mutable_data(), shape(), strides(), ndim());
+    }
+
+    /**
+     * Returns a proxy object that provides const access to the array's data without bounds or
+     * dimensionality checking.  Unlike `mutable_unchecked()`, this does not require that the
+     * underlying array have the `writable` flag.  Use with care: the array must not be destroyed
+     * or reshaped for the duration of the returned object, and the caller must take care not to
+     * access invalid dimensions or dimension indices.
+     */
+    template <typename T, ssize_t Dims = -1>
+    detail::unchecked_reference<T, Dims> unchecked() const & {
+        if (Dims >= 0 && ndim() != Dims) {
+            throw std::domain_error("array has incorrect number of dimensions: "
+                                    + std::to_string(ndim()) + "; expected "
+                                    + std::to_string(Dims));
+        }
+        return detail::unchecked_reference<T, Dims>(data(), shape(), strides(), ndim());
+    }
+
+    /// Return a new view with all of the dimensions of length 1 removed
+    array squeeze() {
+        auto &api = detail::npy_api::get();
+        return reinterpret_steal<array>(api.PyArray_Squeeze_(m_ptr));
+    }
+
+    /// Resize array to given shape
+    /// If refcheck is true and more that one reference exist to this array
+    /// then resize will succeed only if it makes a reshape, i.e. original size doesn't change
+    void resize(ShapeContainer new_shape, bool refcheck = true) {
+        detail::npy_api::PyArray_Dims d
+            = {// Use reinterpret_cast for PyPy on Windows (remove if fixed, checked on 7.3.1)
+               reinterpret_cast<Py_intptr_t *>(new_shape->data()),
+               int(new_shape->size())};
+        // try to resize, set ordering param to -1 cause it's not used anyway
+        auto new_array = reinterpret_steal<object>(
+            detail::npy_api::get().PyArray_Resize_(m_ptr, &d, int(refcheck), -1));
+        if (!new_array) {
+            throw error_already_set();
+        }
+        if (isinstance<array>(new_array)) {
+            *this = std::move(new_array);
+        }
+    }
+
+    /// Optional `order` parameter omitted, to be added as needed.
+    array reshape(ShapeContainer new_shape) {
+        detail::npy_api::PyArray_Dims d
+            = {reinterpret_cast<Py_intptr_t *>(new_shape->data()), int(new_shape->size())};
+        auto new_array
+            = reinterpret_steal<array>(detail::npy_api::get().PyArray_Newshape_(m_ptr, &d, 0));
+        if (!new_array) {
+            throw error_already_set();
+        }
+        return new_array;
+    }
+
+    /// Create a view of an array in a different data type.
+    /// This function may fundamentally reinterpret the data in the array.
+    /// It is the responsibility of the caller to ensure that this is safe.
+    /// Only supports the `dtype` argument, the `type` argument is omitted,
+    /// to be added as needed.
+    array view(const std::string &dtype) {
+        auto &api = detail::npy_api::get();
+        auto new_view = reinterpret_steal<array>(api.PyArray_View_(
+            m_ptr, dtype::from_args(pybind11::str(dtype)).release().ptr(), nullptr));
+        if (!new_view) {
+            throw error_already_set();
+        }
+        return new_view;
+    }
+
+    /// Ensure that the argument is a NumPy array
+    /// In case of an error, nullptr is returned and the Python error is cleared.
+    static array ensure(handle h, int ExtraFlags = 0) {
+        auto result = reinterpret_steal<array>(raw_array(h.ptr(), ExtraFlags));
+        if (!result) {
+            PyErr_Clear();
+        }
+        return result;
+    }
+
+protected:
+    template <typename, typename>
+    friend struct detail::npy_format_descriptor;
+
+    void fail_dim_check(ssize_t dim, const std::string &msg) const {
+        throw index_error(msg + ": " + std::to_string(dim) + " (ndim = " + std::to_string(ndim())
+                          + ')');
+    }
+
+    template <typename... Ix>
+    ssize_t byte_offset(Ix... index) const {
+        check_dimensions(index...);
+        return detail::byte_offset_unsafe(strides(), ssize_t(index)...);
+    }
+
+    void check_writeable() const {
+        if (!writeable()) {
+            throw std::domain_error("array is not writeable");
+        }
+    }
+
+    template <typename... Ix>
+    void check_dimensions(Ix... index) const {
+        check_dimensions_impl(ssize_t(0), shape(), ssize_t(index)...);
+    }
+
+    void check_dimensions_impl(ssize_t, const ssize_t *) const {}
+
+    template <typename... Ix>
+    void check_dimensions_impl(ssize_t axis, const ssize_t *shape, ssize_t i, Ix... index) const {
+        if (i >= *shape) {
+            throw index_error(std::string("index ") + std::to_string(i)
+                              + " is out of bounds for axis " + std::to_string(axis)
+                              + " with size " + std::to_string(*shape));
+        }
+        check_dimensions_impl(axis + 1, shape + 1, index...);
+    }
+
+    /// Create array from any object -- always returns a new reference
+    static PyObject *raw_array(PyObject *ptr, int ExtraFlags = 0) {
+        if (ptr == nullptr) {
+            set_error(PyExc_ValueError, "cannot create a pybind11::array from a nullptr");
+            return nullptr;
+        }
+        return detail::npy_api::get().PyArray_FromAny_(
+            ptr, nullptr, 0, 0, detail::npy_api::NPY_ARRAY_ENSUREARRAY_ | ExtraFlags, nullptr);
+    }
+};
+
+template <typename T, int ExtraFlags = array::forcecast>
+class array_t : public array {
+private:
+    struct private_ctor {};
+    // Delegating constructor needed when both moving and accessing in the same constructor
+    array_t(private_ctor,
+            ShapeContainer &&shape,
+            StridesContainer &&strides,
+            const T *ptr,
+            handle base)
+        : array(std::move(shape), std::move(strides), ptr, base) {}
+
+public:
+    static_assert(!detail::array_info<T>::is_array, "Array types cannot be used with array_t");
+
+    using value_type = T;
+
+    array_t() : array(0, static_cast<const T *>(nullptr)) {}
+    array_t(handle h, borrowed_t) : array(h, borrowed_t{}) {}
+    array_t(handle h, stolen_t) : array(h, stolen_t{}) {}
+
+    PYBIND11_DEPRECATED("Use array_t<T>::ensure() instead")
+    array_t(handle h, bool is_borrowed) : array(raw_array_t(h.ptr()), stolen_t{}) {
+        if (!m_ptr) {
+            PyErr_Clear();
+        }
+        if (!is_borrowed) {
+            Py_XDECREF(h.ptr());
+        }
+    }
+
+    // NOLINTNEXTLINE(google-explicit-constructor)
+    array_t(const object &o) : array(raw_array_t(o.ptr()), stolen_t{}) {
+        if (!m_ptr) {
+            throw error_already_set();
+        }
+    }
+
+    explicit array_t(const buffer_info &info, handle base = handle()) : array(info, base) {}
+
+    array_t(ShapeContainer shape,
+            StridesContainer strides,
+            const T *ptr = nullptr,
+            handle base = handle())
+        : array(std::move(shape), std::move(strides), ptr, base) {}
+
+    explicit array_t(ShapeContainer shape, const T *ptr = nullptr, handle base = handle())
+        : array_t(private_ctor{},
+                  std::move(shape),
+                  (ExtraFlags & f_style) != 0 ? detail::f_strides(*shape, itemsize())
+                                              : detail::c_strides(*shape, itemsize()),
+                  ptr,
+                  base) {}
+
+    explicit array_t(ssize_t count, const T *ptr = nullptr, handle base = handle())
+        : array({count}, {}, ptr, base) {}
+
+    constexpr ssize_t itemsize() const { return sizeof(T); }
+
+    template <typename... Ix>
+    ssize_t index_at(Ix... index) const {
+        return offset_at(index...) / itemsize();
+    }
+
+    template <typename... Ix>
+    const T *data(Ix... index) const {
+        return static_cast<const T *>(array::data(index...));
+    }
+
+    template <typename... Ix>
+    T *mutable_data(Ix... index) {
+        return static_cast<T *>(array::mutable_data(index...));
+    }
+
+    // Reference to element at a given index
+    template <typename... Ix>
+    const T &at(Ix... index) const {
+        if ((ssize_t) sizeof...(index) != ndim()) {
+            fail_dim_check(sizeof...(index), "index dimension mismatch");
+        }
+        return *(static_cast<const T *>(array::data())
+                 + byte_offset(ssize_t(index)...) / itemsize());
+    }
+
+    // Mutable reference to element at a given index
+    template <typename... Ix>
+    T &mutable_at(Ix... index) {
+        if ((ssize_t) sizeof...(index) != ndim()) {
+            fail_dim_check(sizeof...(index), "index dimension mismatch");
+        }
+        return *(static_cast<T *>(array::mutable_data())
+                 + byte_offset(ssize_t(index)...) / itemsize());
+    }
+
+    /**
+     * Returns a proxy object that provides access to the array's data without bounds or
+     * dimensionality checking.  Will throw if the array is missing the `writeable` flag.  Use with
+     * care: the array must not be destroyed or reshaped for the duration of the returned object,
+     * and the caller must take care not to access invalid dimensions or dimension indices.
+     */
+    template <ssize_t Dims = -1>
+    detail::unchecked_mutable_reference<T, Dims> mutable_unchecked() & {
+        return array::mutable_unchecked<T, Dims>();
+    }
+
+    /**
+     * Returns a proxy object that provides const access to the array's data without bounds or
+     * dimensionality checking.  Unlike `mutable_unchecked()`, this does not require that the
+     * underlying array have the `writable` flag.  Use with care: the array must not be destroyed
+     * or reshaped for the duration of the returned object, and the caller must take care not to
+     * access invalid dimensions or dimension indices.
+     */
+    template <ssize_t Dims = -1>
+    detail::unchecked_reference<T, Dims> unchecked() const & {
+        return array::unchecked<T, Dims>();
+    }
+
+    /// Ensure that the argument is a NumPy array of the correct dtype (and if not, try to convert
+    /// it).  In case of an error, nullptr is returned and the Python error is cleared.
+    static array_t ensure(handle h) {
+        auto result = reinterpret_steal<array_t>(raw_array_t(h.ptr()));
+        if (!result) {
+            PyErr_Clear();
+        }
+        return result;
+    }
+
+    static bool check_(handle h) {
+        const auto &api = detail::npy_api::get();
+        return api.PyArray_Check_(h.ptr())
+               && api.PyArray_EquivTypes_(detail::array_proxy(h.ptr())->descr,
+                                          dtype::of<T>().ptr())
+               && detail::check_flags(h.ptr(), ExtraFlags & (array::c_style | array::f_style));
+    }
+
+protected:
+    /// Create array from any object -- always returns a new reference
+    static PyObject *raw_array_t(PyObject *ptr) {
+        if (ptr == nullptr) {
+            set_error(PyExc_ValueError, "cannot create a pybind11::array_t from a nullptr");
+            return nullptr;
+        }
+        return detail::npy_api::get().PyArray_FromAny_(ptr,
+                                                       dtype::of<T>().release().ptr(),
+                                                       0,
+                                                       0,
+                                                       detail::npy_api::NPY_ARRAY_ENSUREARRAY_
+                                                           | ExtraFlags,
+                                                       nullptr);
+    }
+};
+
+template <typename T>
+struct format_descriptor<T, detail::enable_if_t<detail::is_pod_struct<T>::value>> {
+    static std::string format() {
+        return detail::npy_format_descriptor<typename std::remove_cv<T>::type>::format();
+    }
+};
+
+template <size_t N>
+struct format_descriptor<char[N]> {
+    static std::string format() { return std::to_string(N) + 's'; }
+};
+template <size_t N>
+struct format_descriptor<std::array<char, N>> {
+    static std::string format() { return std::to_string(N) + 's'; }
+};
+
+template <typename T>
+struct format_descriptor<T, detail::enable_if_t<std::is_enum<T>::value>> {
+    static std::string format() {
+        return format_descriptor<
+            typename std::remove_cv<typename std::underlying_type<T>::type>::type>::format();
+    }
+};
+
+template <typename T>
+struct format_descriptor<T, detail::enable_if_t<detail::array_info<T>::is_array>> {
+    static std::string format() {
+        using namespace detail;
+        static constexpr auto extents = const_name("(") + array_info<T>::extents + const_name(")");
+        return extents.text + format_descriptor<remove_all_extents_t<T>>::format();
+    }
+};
+
+PYBIND11_NAMESPACE_BEGIN(detail)
+template <typename T, int ExtraFlags>
+struct pyobject_caster<array_t<T, ExtraFlags>> {
+    using type = array_t<T, ExtraFlags>;
+
+    bool load(handle src, bool convert) {
+        if (!convert && !type::check_(src)) {
+            return false;
+        }
+        value = type::ensure(src);
+        return static_cast<bool>(value);
+    }
+
+    static handle cast(const handle &src, return_value_policy /* policy */, handle /* parent */) {
+        return src.inc_ref();
+    }
+    PYBIND11_TYPE_CASTER(type, handle_type_name<type>::name);
+};
+
+template <typename T>
+struct compare_buffer_info<T, detail::enable_if_t<detail::is_pod_struct<T>::value>> {
+    static bool compare(const buffer_info &b) {
+        return npy_api::get().PyArray_EquivTypes_(dtype::of<T>().ptr(), dtype(b).ptr());
+    }
+};
+
+template <typename T, typename = void>
+struct npy_format_descriptor_name;
+
+template <typename T>
+struct npy_format_descriptor_name<T, enable_if_t<std::is_integral<T>::value>> {
+    static constexpr auto name = const_name<std::is_same<T, bool>::value>(
+        const_name("bool"),
+        const_name<std::is_signed<T>::value>("numpy.int", "numpy.uint")
+            + const_name<sizeof(T) * 8>());
+};
+
+template <typename T>
+struct npy_format_descriptor_name<T, enable_if_t<std::is_floating_point<T>::value>> {
+    static constexpr auto name = const_name < std::is_same<T, float>::value
+                                 || std::is_same<T, const float>::value
+                                 || std::is_same<T, double>::value
+                                 || std::is_same<T, const double>::value
+                                        > (const_name("numpy.float") + const_name<sizeof(T) * 8>(),
+                                           const_name("numpy.longdouble"));
+};
+
+template <typename T>
+struct npy_format_descriptor_name<T, enable_if_t<is_complex<T>::value>> {
+    static constexpr auto name = const_name < std::is_same<typename T::value_type, float>::value
+                                 || std::is_same<typename T::value_type, const float>::value
+                                 || std::is_same<typename T::value_type, double>::value
+                                 || std::is_same<typename T::value_type, const double>::value
+                                        > (const_name("numpy.complex")
+                                               + const_name<sizeof(typename T::value_type) * 16>(),
+                                           const_name("numpy.longcomplex"));
+};
+
+template <typename T>
+struct npy_format_descriptor<
+    T,
+    enable_if_t<satisfies_any_of<T, std::is_arithmetic, is_complex>::value>>
+    : npy_format_descriptor_name<T> {
+private:
+    // NB: the order here must match the one in common.h
+    constexpr static const int values[15] = {npy_api::NPY_BOOL_,
+                                             npy_api::NPY_BYTE_,
+                                             npy_api::NPY_UBYTE_,
+                                             npy_api::NPY_INT16_,
+                                             npy_api::NPY_UINT16_,
+                                             npy_api::NPY_INT32_,
+                                             npy_api::NPY_UINT32_,
+                                             npy_api::NPY_INT64_,
+                                             npy_api::NPY_UINT64_,
+                                             npy_api::NPY_FLOAT_,
+                                             npy_api::NPY_DOUBLE_,
+                                             npy_api::NPY_LONGDOUBLE_,
+                                             npy_api::NPY_CFLOAT_,
+                                             npy_api::NPY_CDOUBLE_,
+                                             npy_api::NPY_CLONGDOUBLE_};
+
+public:
+    static constexpr int value = values[detail::is_fmt_numeric<T>::index];
+
+    static pybind11::dtype dtype() { return pybind11::dtype(/*typenum*/ value); }
+};
+
+template <typename T>
+struct npy_format_descriptor<T, enable_if_t<is_same_ignoring_cvref<T, PyObject *>::value>> {
+    static constexpr auto name = const_name("object");
+
+    static constexpr int value = npy_api::NPY_OBJECT_;
+
+    static pybind11::dtype dtype() { return pybind11::dtype(/*typenum*/ value); }
+};
+
+#define PYBIND11_DECL_CHAR_FMT                                                                    \
+    static constexpr auto name = const_name("S") + const_name<N>();                               \
+    static pybind11::dtype dtype() {                                                              \
+        return pybind11::dtype(std::string("S") + std::to_string(N));                             \
+    }
+template <size_t N>
+struct npy_format_descriptor<char[N]> {
+    PYBIND11_DECL_CHAR_FMT
+};
+template <size_t N>
+struct npy_format_descriptor<std::array<char, N>> {
+    PYBIND11_DECL_CHAR_FMT
+};
+#undef PYBIND11_DECL_CHAR_FMT
+
+template <typename T>
+struct npy_format_descriptor<T, enable_if_t<array_info<T>::is_array>> {
+private:
+    using base_descr = npy_format_descriptor<typename array_info<T>::type>;
+
+public:
+    static_assert(!array_info<T>::is_empty, "Zero-sized arrays are not supported");
+
+    static constexpr auto name
+        = const_name("(") + array_info<T>::extents + const_name(")") + base_descr::name;
+    static pybind11::dtype dtype() {
+        list shape;
+        array_info<T>::append_extents(shape);
+        return pybind11::dtype::from_args(
+            pybind11::make_tuple(base_descr::dtype(), std::move(shape)));
+    }
+};
+
+template <typename T>
+struct npy_format_descriptor<T, enable_if_t<std::is_enum<T>::value>> {
+private:
+    using base_descr = npy_format_descriptor<typename std::underlying_type<T>::type>;
+
+public:
+    static constexpr auto name = base_descr::name;
+    static pybind11::dtype dtype() { return base_descr::dtype(); }
+};
+
+struct field_descriptor {
+    const char *name;
+    ssize_t offset;
+    ssize_t size;
+    std::string format;
+    dtype descr;
+};
+
+PYBIND11_NOINLINE void register_structured_dtype(any_container<field_descriptor> fields,
+                                                 const std::type_info &tinfo,
+                                                 ssize_t itemsize,
+                                                 bool (*direct_converter)(PyObject *, void *&)) {
+
+    auto &numpy_internals = get_numpy_internals();
+    if (numpy_internals.get_type_info(tinfo, false)) {
+        pybind11_fail("NumPy: dtype is already registered");
+    }
+
+    // Use ordered fields because order matters as of NumPy 1.14:
+    // https://docs.scipy.org/doc/numpy/release.html#multiple-field-indexing-assignment-of-structured-arrays
+    std::vector<field_descriptor> ordered_fields(std::move(fields));
+    std::sort(
+        ordered_fields.begin(),
+        ordered_fields.end(),
+        [](const field_descriptor &a, const field_descriptor &b) { return a.offset < b.offset; });
+
+    list names, formats, offsets;
+    for (auto &field : ordered_fields) {
+        if (!field.descr) {
+            pybind11_fail(std::string("NumPy: unsupported field dtype: `") + field.name + "` @ "
+                          + tinfo.name());
+        }
+        names.append(pybind11::str(field.name));
+        formats.append(field.descr);
+        offsets.append(pybind11::int_(field.offset));
+    }
+    auto *dtype_ptr
+        = pybind11::dtype(std::move(names), std::move(formats), std::move(offsets), itemsize)
+              .release()
+              .ptr();
+
+    // There is an existing bug in NumPy (as of v1.11): trailing bytes are
+    // not encoded explicitly into the format string. This will supposedly
+    // get fixed in v1.12; for further details, see these:
+    // - https://github.com/numpy/numpy/issues/7797
+    // - https://github.com/numpy/numpy/pull/7798
+    // Because of this, we won't use numpy's logic to generate buffer format
+    // strings and will just do it ourselves.
+    ssize_t offset = 0;
+    std::ostringstream oss;
+    // mark the structure as unaligned with '^', because numpy and C++ don't
+    // always agree about alignment (particularly for complex), and we're
+    // explicitly listing all our padding. This depends on none of the fields
+    // overriding the endianness. Putting the ^ in front of individual fields
+    // isn't guaranteed to work due to https://github.com/numpy/numpy/issues/9049
+    oss << "^T{";
+    for (auto &field : ordered_fields) {
+        if (field.offset > offset) {
+            oss << (field.offset - offset) << 'x';
+        }
+        oss << field.format << ':' << field.name << ':';
+        offset = field.offset + field.size;
+    }
+    if (itemsize > offset) {
+        oss << (itemsize - offset) << 'x';
+    }
+    oss << '}';
+    auto format_str = oss.str();
+
+    // Smoke test: verify that NumPy properly parses our buffer format string
+    auto &api = npy_api::get();
+    auto arr = array(buffer_info(nullptr, itemsize, format_str, 1));
+    if (!api.PyArray_EquivTypes_(dtype_ptr, arr.dtype().ptr())) {
+        pybind11_fail("NumPy: invalid buffer descriptor!");
+    }
+
+    auto tindex = std::type_index(tinfo);
+    numpy_internals.registered_dtypes[tindex] = {dtype_ptr, std::move(format_str)};
+    with_internals([tindex, &direct_converter](internals &internals) {
+        internals.direct_conversions[tindex].push_back(direct_converter);
+    });
+}
+
+template <typename T, typename SFINAE>
+struct npy_format_descriptor {
+    static_assert(is_pod_struct<T>::value,
+                  "Attempt to use a non-POD or unimplemented POD type as a numpy dtype");
+
+    static constexpr auto name = make_caster<T>::name;
+
+    static pybind11::dtype dtype() { return reinterpret_borrow<pybind11::dtype>(dtype_ptr()); }
+
+    static std::string format() {
+        static auto format_str = get_numpy_internals().get_type_info<T>(true)->format_str;
+        return format_str;
+    }
+
+    static void register_dtype(any_container<field_descriptor> fields) {
+        register_structured_dtype(std::move(fields),
+                                  typeid(typename std::remove_cv<T>::type),
+                                  sizeof(T),
+                                  &direct_converter);
+    }
+
+private:
+    static PyObject *dtype_ptr() {
+        static PyObject *ptr = get_numpy_internals().get_type_info<T>(true)->dtype_ptr;
+        return ptr;
+    }
+
+    static bool direct_converter(PyObject *obj, void *&value) {
+        auto &api = npy_api::get();
+        if (!PyObject_TypeCheck(obj, api.PyVoidArrType_Type_)) {
+            return false;
+        }
+        if (auto descr = reinterpret_steal<object>(api.PyArray_DescrFromScalar_(obj))) {
+            if (api.PyArray_EquivTypes_(dtype_ptr(), descr.ptr())) {
+                value = ((PyVoidScalarObject_Proxy *) obj)->obval;
+                return true;
+            }
+        }
+        return false;
+    }
+};
+
+#ifdef __CLION_IDE__ // replace heavy macro with dummy code for the IDE (doesn't affect code)
+#    define PYBIND11_NUMPY_DTYPE(Type, ...) ((void) 0)
+#    define PYBIND11_NUMPY_DTYPE_EX(Type, ...) ((void) 0)
+#else
+
+#    define PYBIND11_FIELD_DESCRIPTOR_EX(T, Field, Name)                                          \
+        ::pybind11::detail::field_descriptor {                                                    \
+            Name, offsetof(T, Field), sizeof(decltype(std::declval<T>().Field)),                  \
+                ::pybind11::format_descriptor<decltype(std::declval<T>().Field)>::format(),       \
+                ::pybind11::detail::npy_format_descriptor<                                        \
+                    decltype(std::declval<T>().Field)>::dtype()                                   \
+        }
+
+// Extract name, offset and format descriptor for a struct field
+#    define PYBIND11_FIELD_DESCRIPTOR(T, Field) PYBIND11_FIELD_DESCRIPTOR_EX(T, Field, #Field)
+
+// The main idea of this macro is borrowed from https://github.com/swansontec/map-macro
+// (C) William Swanson, Paul Fultz
+#    define PYBIND11_EVAL0(...) __VA_ARGS__
+#    define PYBIND11_EVAL1(...) PYBIND11_EVAL0(PYBIND11_EVAL0(PYBIND11_EVAL0(__VA_ARGS__)))
+#    define PYBIND11_EVAL2(...) PYBIND11_EVAL1(PYBIND11_EVAL1(PYBIND11_EVAL1(__VA_ARGS__)))
+#    define PYBIND11_EVAL3(...) PYBIND11_EVAL2(PYBIND11_EVAL2(PYBIND11_EVAL2(__VA_ARGS__)))
+#    define PYBIND11_EVAL4(...) PYBIND11_EVAL3(PYBIND11_EVAL3(PYBIND11_EVAL3(__VA_ARGS__)))
+#    define PYBIND11_EVAL(...) PYBIND11_EVAL4(PYBIND11_EVAL4(PYBIND11_EVAL4(__VA_ARGS__)))
+#    define PYBIND11_MAP_END(...)
+#    define PYBIND11_MAP_OUT
+#    define PYBIND11_MAP_COMMA ,
+#    define PYBIND11_MAP_GET_END() 0, PYBIND11_MAP_END
+#    define PYBIND11_MAP_NEXT0(test, next, ...) next PYBIND11_MAP_OUT
+#    define PYBIND11_MAP_NEXT1(test, next) PYBIND11_MAP_NEXT0(test, next, 0)
+#    define PYBIND11_MAP_NEXT(test, next) PYBIND11_MAP_NEXT1(PYBIND11_MAP_GET_END test, next)
+#    if defined(_MSC_VER)                                                                         \
+        && !defined(__clang__) // MSVC is not as eager to expand macros, hence this workaround
+#        define PYBIND11_MAP_LIST_NEXT1(test, next)                                               \
+            PYBIND11_EVAL0(PYBIND11_MAP_NEXT0(test, PYBIND11_MAP_COMMA next, 0))
+#    else
+#        define PYBIND11_MAP_LIST_NEXT1(test, next)                                               \
+            PYBIND11_MAP_NEXT0(test, PYBIND11_MAP_COMMA next, 0)
+#    endif
+#    define PYBIND11_MAP_LIST_NEXT(test, next)                                                    \
+        PYBIND11_MAP_LIST_NEXT1(PYBIND11_MAP_GET_END test, next)
+#    define PYBIND11_MAP_LIST0(f, t, x, peek, ...)                                                \
+        f(t, x) PYBIND11_MAP_LIST_NEXT(peek, PYBIND11_MAP_LIST1)(f, t, peek, __VA_ARGS__)
+#    define PYBIND11_MAP_LIST1(f, t, x, peek, ...)                                                \
+        f(t, x) PYBIND11_MAP_LIST_NEXT(peek, PYBIND11_MAP_LIST0)(f, t, peek, __VA_ARGS__)
+// PYBIND11_MAP_LIST(f, t, a1, a2, ...) expands to f(t, a1), f(t, a2), ...
+#    define PYBIND11_MAP_LIST(f, t, ...)                                                          \
+        PYBIND11_EVAL(PYBIND11_MAP_LIST1(f, t, __VA_ARGS__, (), 0))
+
+#    define PYBIND11_NUMPY_DTYPE(Type, ...)                                                       \
+        ::pybind11::detail::npy_format_descriptor<Type>::register_dtype(                          \
+            ::std::vector<::pybind11::detail::field_descriptor>{                                  \
+                PYBIND11_MAP_LIST(PYBIND11_FIELD_DESCRIPTOR, Type, __VA_ARGS__)})
+
+#    if defined(_MSC_VER) && !defined(__clang__)
+#        define PYBIND11_MAP2_LIST_NEXT1(test, next)                                              \
+            PYBIND11_EVAL0(PYBIND11_MAP_NEXT0(test, PYBIND11_MAP_COMMA next, 0))
+#    else
+#        define PYBIND11_MAP2_LIST_NEXT1(test, next)                                              \
+            PYBIND11_MAP_NEXT0(test, PYBIND11_MAP_COMMA next, 0)
+#    endif
+#    define PYBIND11_MAP2_LIST_NEXT(test, next)                                                   \
+        PYBIND11_MAP2_LIST_NEXT1(PYBIND11_MAP_GET_END test, next)
+#    define PYBIND11_MAP2_LIST0(f, t, x1, x2, peek, ...)                                          \
+        f(t, x1, x2) PYBIND11_MAP2_LIST_NEXT(peek, PYBIND11_MAP2_LIST1)(f, t, peek, __VA_ARGS__)
+#    define PYBIND11_MAP2_LIST1(f, t, x1, x2, peek, ...)                                          \
+        f(t, x1, x2) PYBIND11_MAP2_LIST_NEXT(peek, PYBIND11_MAP2_LIST0)(f, t, peek, __VA_ARGS__)
+// PYBIND11_MAP2_LIST(f, t, a1, a2, ...) expands to f(t, a1, a2), f(t, a3, a4), ...
+#    define PYBIND11_MAP2_LIST(f, t, ...)                                                         \
+        PYBIND11_EVAL(PYBIND11_MAP2_LIST1(f, t, __VA_ARGS__, (), 0))
+
+#    define PYBIND11_NUMPY_DTYPE_EX(Type, ...)                                                    \
+        ::pybind11::detail::npy_format_descriptor<Type>::register_dtype(                          \
+            ::std::vector<::pybind11::detail::field_descriptor>{                                  \
+                PYBIND11_MAP2_LIST(PYBIND11_FIELD_DESCRIPTOR_EX, Type, __VA_ARGS__)})
+
+#endif // __CLION_IDE__
+
+class common_iterator {
+public:
+    using container_type = std::vector<ssize_t>;
+    using value_type = container_type::value_type;
+    using size_type = container_type::size_type;
+
+    common_iterator() : m_strides() {}
+
+    common_iterator(void *ptr, const container_type &strides, const container_type &shape)
+        : p_ptr(reinterpret_cast<char *>(ptr)), m_strides(strides.size()) {
+        m_strides.back() = static_cast<value_type>(strides.back());
+        for (size_type i = m_strides.size() - 1; i != 0; --i) {
+            size_type j = i - 1;
+            auto s = static_cast<value_type>(shape[i]);
+            m_strides[j] = strides[j] + m_strides[i] - strides[i] * s;
+        }
+    }
+
+    void increment(size_type dim) { p_ptr += m_strides[dim]; }
+
+    void *data() const { return p_ptr; }
+
+private:
+    char *p_ptr{nullptr};
+    container_type m_strides;
+};
+
+template <size_t N>
+class multi_array_iterator {
+public:
+    using container_type = std::vector<ssize_t>;
+
+    multi_array_iterator(const std::array<buffer_info, N> &buffers, const container_type &shape)
+        : m_shape(shape.size()), m_index(shape.size(), 0), m_common_iterator() {
+
+        // Manual copy to avoid conversion warning if using std::copy
+        for (size_t i = 0; i < shape.size(); ++i) {
+            m_shape[i] = shape[i];
+        }
+
+        container_type strides(shape.size());
+        for (size_t i = 0; i < N; ++i) {
+            init_common_iterator(buffers[i], shape, m_common_iterator[i], strides);
+        }
+    }
+
+    multi_array_iterator &operator++() {
+        for (size_t j = m_index.size(); j != 0; --j) {
+            size_t i = j - 1;
+            if (++m_index[i] != m_shape[i]) {
+                increment_common_iterator(i);
+                break;
+            }
+            m_index[i] = 0;
+        }
+        return *this;
+    }
+
+    template <size_t K, class T = void>
+    T *data() const {
+        return reinterpret_cast<T *>(m_common_iterator[K].data());
+    }
+
+private:
+    using common_iter = common_iterator;
+
+    void init_common_iterator(const buffer_info &buffer,
+                              const container_type &shape,
+                              common_iter &iterator,
+                              container_type &strides) {
+        auto buffer_shape_iter = buffer.shape.rbegin();
+        auto buffer_strides_iter = buffer.strides.rbegin();
+        auto shape_iter = shape.rbegin();
+        auto strides_iter = strides.rbegin();
+
+        while (buffer_shape_iter != buffer.shape.rend()) {
+            if (*shape_iter == *buffer_shape_iter) {
+                *strides_iter = *buffer_strides_iter;
+            } else {
+                *strides_iter = 0;
+            }
+
+            ++buffer_shape_iter;
+            ++buffer_strides_iter;
+            ++shape_iter;
+            ++strides_iter;
+        }
+
+        std::fill(strides_iter, strides.rend(), 0);
+        iterator = common_iter(buffer.ptr, strides, shape);
+    }
+
+    void increment_common_iterator(size_t dim) {
+        for (auto &iter : m_common_iterator) {
+            iter.increment(dim);
+        }
+    }
+
+    container_type m_shape;
+    container_type m_index;
+    std::array<common_iter, N> m_common_iterator;
+};
+
+enum class broadcast_trivial { non_trivial, c_trivial, f_trivial };
+
+// Populates the shape and number of dimensions for the set of buffers.  Returns a
+// broadcast_trivial enum value indicating whether the broadcast is "trivial"--that is, has each
+// buffer being either a singleton or a full-size, C-contiguous (`c_trivial`) or Fortran-contiguous
+// (`f_trivial`) storage buffer; returns `non_trivial` otherwise.
+template <size_t N>
+broadcast_trivial
+broadcast(const std::array<buffer_info, N> &buffers, ssize_t &ndim, std::vector<ssize_t> &shape) {
+    ndim = std::accumulate(
+        buffers.begin(), buffers.end(), ssize_t(0), [](ssize_t res, const buffer_info &buf) {
+            return std::max(res, buf.ndim);
+        });
+
+    shape.clear();
+    shape.resize((size_t) ndim, 1);
+
+    // Figure out the output size, and make sure all input arrays conform (i.e. are either size 1
+    // or the full size).
+    for (size_t i = 0; i < N; ++i) {
+        auto res_iter = shape.rbegin();
+        auto end = buffers[i].shape.rend();
+        for (auto shape_iter = buffers[i].shape.rbegin(); shape_iter != end;
+             ++shape_iter, ++res_iter) {
+            const auto &dim_size_in = *shape_iter;
+            auto &dim_size_out = *res_iter;
+
+            // Each input dimension can either be 1 or `n`, but `n` values must match across
+            // buffers
+            if (dim_size_out == 1) {
+                dim_size_out = dim_size_in;
+            } else if (dim_size_in != 1 && dim_size_in != dim_size_out) {
+                pybind11_fail("pybind11::vectorize: incompatible size/dimension of inputs!");
+            }
+        }
+    }
+
+    bool trivial_broadcast_c = true;
+    bool trivial_broadcast_f = true;
+    for (size_t i = 0; i < N && (trivial_broadcast_c || trivial_broadcast_f); ++i) {
+        if (buffers[i].size == 1) {
+            continue;
+        }
+
+        // Require the same number of dimensions:
+        if (buffers[i].ndim != ndim) {
+            return broadcast_trivial::non_trivial;
+        }
+
+        // Require all dimensions be full-size:
+        if (!std::equal(buffers[i].shape.cbegin(), buffers[i].shape.cend(), shape.cbegin())) {
+            return broadcast_trivial::non_trivial;
+        }
+
+        // Check for C contiguity (but only if previous inputs were also C contiguous)
+        if (trivial_broadcast_c) {
+            ssize_t expect_stride = buffers[i].itemsize;
+            auto end = buffers[i].shape.crend();
+            for (auto shape_iter = buffers[i].shape.crbegin(),
+                      stride_iter = buffers[i].strides.crbegin();
+                 trivial_broadcast_c && shape_iter != end;
+                 ++shape_iter, ++stride_iter) {
+                if (expect_stride == *stride_iter) {
+                    expect_stride *= *shape_iter;
+                } else {
+                    trivial_broadcast_c = false;
+                }
+            }
+        }
+
+        // Check for Fortran contiguity (if previous inputs were also F contiguous)
+        if (trivial_broadcast_f) {
+            ssize_t expect_stride = buffers[i].itemsize;
+            auto end = buffers[i].shape.cend();
+            for (auto shape_iter = buffers[i].shape.cbegin(),
+                      stride_iter = buffers[i].strides.cbegin();
+                 trivial_broadcast_f && shape_iter != end;
+                 ++shape_iter, ++stride_iter) {
+                if (expect_stride == *stride_iter) {
+                    expect_stride *= *shape_iter;
+                } else {
+                    trivial_broadcast_f = false;
+                }
+            }
+        }
+    }
+
+    return trivial_broadcast_c   ? broadcast_trivial::c_trivial
+           : trivial_broadcast_f ? broadcast_trivial::f_trivial
+                                 : broadcast_trivial::non_trivial;
+}
+
+template <typename T>
+struct vectorize_arg {
+    static_assert(!std::is_rvalue_reference<T>::value,
+                  "Functions with rvalue reference arguments cannot be vectorized");
+    // The wrapped function gets called with this type:
+    using call_type = remove_reference_t<T>;
+    // Is this a vectorized argument?
+    static constexpr bool vectorize
+        = satisfies_any_of<call_type, std::is_arithmetic, is_complex, is_pod>::value
+          && satisfies_none_of<call_type,
+                               std::is_pointer,
+                               std::is_array,
+                               is_std_array,
+                               std::is_enum>::value
+          && (!std::is_reference<T>::value
+              || (std::is_lvalue_reference<T>::value && std::is_const<call_type>::value));
+    // Accept this type: an array for vectorized types, otherwise the type as-is:
+    using type = conditional_t<vectorize, array_t<remove_cv_t<call_type>, array::forcecast>, T>;
+};
+
+// py::vectorize when a return type is present
+template <typename Func, typename Return, typename... Args>
+struct vectorize_returned_array {
+    using Type = array_t<Return>;
+
+    static Type create(broadcast_trivial trivial, const std::vector<ssize_t> &shape) {
+        if (trivial == broadcast_trivial::f_trivial) {
+            return array_t<Return, array::f_style>(shape);
+        }
+        return array_t<Return>(shape);
+    }
+
+    static Return *mutable_data(Type &array) { return array.mutable_data(); }
+
+    static Return call(Func &f, Args &...args) { return f(args...); }
+
+    static void call(Return *out, size_t i, Func &f, Args &...args) { out[i] = f(args...); }
+};
+
+// py::vectorize when a return type is not present
+template <typename Func, typename... Args>
+struct vectorize_returned_array<Func, void, Args...> {
+    using Type = none;
+
+    static Type create(broadcast_trivial, const std::vector<ssize_t> &) { return none(); }
+
+    static void *mutable_data(Type &) { return nullptr; }
+
+    static detail::void_type call(Func &f, Args &...args) {
+        f(args...);
+        return {};
+    }
+
+    static void call(void *, size_t, Func &f, Args &...args) { f(args...); }
+};
+
+template <typename Func, typename Return, typename... Args>
+struct vectorize_helper {
+
+// NVCC for some reason breaks if NVectorized is private
+#ifdef __CUDACC__
+public:
+#else
+private:
+#endif
+
+    static constexpr size_t N = sizeof...(Args);
+    static constexpr size_t NVectorized = constexpr_sum(vectorize_arg<Args>::vectorize...);
+    static_assert(
+        NVectorized >= 1,
+        "pybind11::vectorize(...) requires a function with at least one vectorizable argument");
+
+public:
+    template <typename T,
+              // SFINAE to prevent shadowing the copy constructor.
+              typename = detail::enable_if_t<
+                  !std::is_same<vectorize_helper, typename std::decay<T>::type>::value>>
+    explicit vectorize_helper(T &&f) : f(std::forward<T>(f)) {}
+
+    object operator()(typename vectorize_arg<Args>::type... args) {
+        return run(args...,
+                   make_index_sequence<N>(),
+                   select_indices<vectorize_arg<Args>::vectorize...>(),
+                   make_index_sequence<NVectorized>());
+    }
+
+private:
+    remove_reference_t<Func> f;
+
+    // Internal compiler error in MSVC 19.16.27025.1 (Visual Studio 2017 15.9.4), when compiling
+    // with "/permissive-" flag when arg_call_types is manually inlined.
+    using arg_call_types = std::tuple<typename vectorize_arg<Args>::call_type...>;
+    template <size_t Index>
+    using param_n_t = typename std::tuple_element<Index, arg_call_types>::type;
+
+    using returned_array = vectorize_returned_array<Func, Return, Args...>;
+
+    // Runs a vectorized function given arguments tuple and three index sequences:
+    //     - Index is the full set of 0 ... (N-1) argument indices;
+    //     - VIndex is the subset of argument indices with vectorized parameters, letting us access
+    //       vectorized arguments (anything not in this sequence is passed through)
+    //     - BIndex is a incremental sequence (beginning at 0) of the same size as VIndex, so that
+    //       we can store vectorized buffer_infos in an array (argument VIndex has its buffer at
+    //       index BIndex in the array).
+    template <size_t... Index, size_t... VIndex, size_t... BIndex>
+    object run(typename vectorize_arg<Args>::type &...args,
+               index_sequence<Index...> i_seq,
+               index_sequence<VIndex...> vi_seq,
+               index_sequence<BIndex...> bi_seq) {
+
+        // Pointers to values the function was called with; the vectorized ones set here will start
+        // out as array_t<T> pointers, but they will be changed them to T pointers before we make
+        // call the wrapped function.  Non-vectorized pointers are left as-is.
+        std::array<void *, N> params{{reinterpret_cast<void *>(&args)...}};
+
+        // The array of `buffer_info`s of vectorized arguments:
+        std::array<buffer_info, NVectorized> buffers{
+            {reinterpret_cast<array *>(params[VIndex])->request()...}};
+
+        /* Determine dimensions parameters of output array */
+        ssize_t nd = 0;
+        std::vector<ssize_t> shape(0);
+        auto trivial = broadcast(buffers, nd, shape);
+        auto ndim = (size_t) nd;
+
+        size_t size
+            = std::accumulate(shape.begin(), shape.end(), (size_t) 1, std::multiplies<size_t>());
+
+        // If all arguments are 0-dimension arrays (i.e. single values) return a plain value (i.e.
+        // not wrapped in an array).
+        if (size == 1 && ndim == 0) {
+            PYBIND11_EXPAND_SIDE_EFFECTS(params[VIndex] = buffers[BIndex].ptr);
+            return cast(
+                returned_array::call(f, *reinterpret_cast<param_n_t<Index> *>(params[Index])...));
+        }
+
+        auto result = returned_array::create(trivial, shape);
+
+        PYBIND11_WARNING_PUSH
+#ifdef PYBIND11_DETECTED_CLANG_WITH_MISLEADING_CALL_STD_MOVE_EXPLICITLY_WARNING
+        PYBIND11_WARNING_DISABLE_CLANG("-Wreturn-std-move")
+#endif
+
+        if (size == 0) {
+            return result;
+        }
+
+        /* Call the function */
+        auto *mutable_data = returned_array::mutable_data(result);
+        if (trivial == broadcast_trivial::non_trivial) {
+            apply_broadcast(buffers, params, mutable_data, size, shape, i_seq, vi_seq, bi_seq);
+        } else {
+            apply_trivial(buffers, params, mutable_data, size, i_seq, vi_seq, bi_seq);
+        }
+
+        return result;
+        PYBIND11_WARNING_POP
+    }
+
+    template <size_t... Index, size_t... VIndex, size_t... BIndex>
+    void apply_trivial(std::array<buffer_info, NVectorized> &buffers,
+                       std::array<void *, N> &params,
+                       Return *out,
+                       size_t size,
+                       index_sequence<Index...>,
+                       index_sequence<VIndex...>,
+                       index_sequence<BIndex...>) {
+
+        // Initialize an array of mutable byte references and sizes with references set to the
+        // appropriate pointer in `params`; as we iterate, we'll increment each pointer by its size
+        // (except for singletons, which get an increment of 0).
+        std::array<std::pair<unsigned char *&, const size_t>, NVectorized> vecparams{
+            {std::pair<unsigned char *&, const size_t>(
+                reinterpret_cast<unsigned char *&>(params[VIndex] = buffers[BIndex].ptr),
+                buffers[BIndex].size == 1 ? 0 : sizeof(param_n_t<VIndex>))...}};
+
+        for (size_t i = 0; i < size; ++i) {
+            returned_array::call(
+                out, i, f, *reinterpret_cast<param_n_t<Index> *>(params[Index])...);
+            for (auto &x : vecparams) {
+                x.first += x.second;
+            }
+        }
+    }
+
+    template <size_t... Index, size_t... VIndex, size_t... BIndex>
+    void apply_broadcast(std::array<buffer_info, NVectorized> &buffers,
+                         std::array<void *, N> &params,
+                         Return *out,
+                         size_t size,
+                         const std::vector<ssize_t> &output_shape,
+                         index_sequence<Index...>,
+                         index_sequence<VIndex...>,
+                         index_sequence<BIndex...>) {
+
+        multi_array_iterator<NVectorized> input_iter(buffers, output_shape);
+
+        for (size_t i = 0; i < size; ++i, ++input_iter) {
+            PYBIND11_EXPAND_SIDE_EFFECTS((params[VIndex] = input_iter.template data<BIndex>()));
+            returned_array::call(
+                out, i, f, *reinterpret_cast<param_n_t<Index> *>(std::get<Index>(params))...);
+        }
+    }
+};
+
+template <typename Func, typename Return, typename... Args>
+vectorize_helper<Func, Return, Args...> vectorize_extractor(const Func &f, Return (*)(Args...)) {
+    return detail::vectorize_helper<Func, Return, Args...>(f);
+}
+
+template <typename T, int Flags>
+struct handle_type_name<array_t<T, Flags>> {
+    static constexpr auto name
+        = const_name("numpy.ndarray[") + npy_format_descriptor<T>::name + const_name("]");
+};
+
+PYBIND11_NAMESPACE_END(detail)
+
+// Vanilla pointer vectorizer:
+template <typename Return, typename... Args>
+detail::vectorize_helper<Return (*)(Args...), Return, Args...> vectorize(Return (*f)(Args...)) {
+    return detail::vectorize_helper<Return (*)(Args...), Return, Args...>(f);
+}
+
+// lambda vectorizer:
+template <typename Func, detail::enable_if_t<detail::is_lambda<Func>::value, int> = 0>
+auto vectorize(Func &&f)
+    -> decltype(detail::vectorize_extractor(std::forward<Func>(f),
+                                            (detail::function_signature_t<Func> *) nullptr)) {
+    return detail::vectorize_extractor(std::forward<Func>(f),
+                                       (detail::function_signature_t<Func> *) nullptr);
+}
+
+// Vectorize a class method (non-const):
+template <typename Return,
+          typename Class,
+          typename... Args,
+          typename Helper = detail::vectorize_helper<
+              decltype(std::mem_fn(std::declval<Return (Class::*)(Args...)>())),
+              Return,
+              Class *,
+              Args...>>
+Helper vectorize(Return (Class::*f)(Args...)) {
+    return Helper(std::mem_fn(f));
+}
+
+// Vectorize a class method (const):
+template <typename Return,
+          typename Class,
+          typename... Args,
+          typename Helper = detail::vectorize_helper<
+              decltype(std::mem_fn(std::declval<Return (Class::*)(Args...) const>())),
+              Return,
+              const Class *,
+              Args...>>
+Helper vectorize(Return (Class::*f)(Args...) const) {
+    return Helper(std::mem_fn(f));
+}
+
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/operators.h

@@ -0,0 +1,202 @@
+/*
+    pybind11/operator.h: Metatemplates for operator overloading
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "pybind11.h"
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+/// Enumeration with all supported operator types
+enum op_id : int {
+    op_add,
+    op_sub,
+    op_mul,
+    op_div,
+    op_mod,
+    op_divmod,
+    op_pow,
+    op_lshift,
+    op_rshift,
+    op_and,
+    op_xor,
+    op_or,
+    op_neg,
+    op_pos,
+    op_abs,
+    op_invert,
+    op_int,
+    op_long,
+    op_float,
+    op_str,
+    op_cmp,
+    op_gt,
+    op_ge,
+    op_lt,
+    op_le,
+    op_eq,
+    op_ne,
+    op_iadd,
+    op_isub,
+    op_imul,
+    op_idiv,
+    op_imod,
+    op_ilshift,
+    op_irshift,
+    op_iand,
+    op_ixor,
+    op_ior,
+    op_complex,
+    op_bool,
+    op_nonzero,
+    op_repr,
+    op_truediv,
+    op_itruediv,
+    op_hash
+};
+
+enum op_type : int {
+    op_l, /* base type on left */
+    op_r, /* base type on right */
+    op_u  /* unary operator */
+};
+
+struct self_t {};
+static const self_t self = self_t();
+
+/// Type for an unused type slot
+struct undefined_t {};
+
+/// Don't warn about an unused variable
+inline self_t __self() { return self; }
+
+/// base template of operator implementations
+template <op_id, op_type, typename B, typename L, typename R>
+struct op_impl {};
+
+/// Operator implementation generator
+template <op_id id, op_type ot, typename L, typename R>
+struct op_ {
+    static constexpr bool op_enable_if_hook = true;
+    template <typename Class, typename... Extra>
+    void execute(Class &cl, const Extra &...extra) const {
+        using Base = typename Class::type;
+        using L_type = conditional_t<std::is_same<L, self_t>::value, Base, L>;
+        using R_type = conditional_t<std::is_same<R, self_t>::value, Base, R>;
+        using op = op_impl<id, ot, Base, L_type, R_type>;
+        cl.def(op::name(), &op::execute, is_operator(), extra...);
+    }
+    template <typename Class, typename... Extra>
+    void execute_cast(Class &cl, const Extra &...extra) const {
+        using Base = typename Class::type;
+        using L_type = conditional_t<std::is_same<L, self_t>::value, Base, L>;
+        using R_type = conditional_t<std::is_same<R, self_t>::value, Base, R>;
+        using op = op_impl<id, ot, Base, L_type, R_type>;
+        cl.def(op::name(), &op::execute_cast, is_operator(), extra...);
+    }
+};
+
+#define PYBIND11_BINARY_OPERATOR(id, rid, op, expr)                                               \
+    template <typename B, typename L, typename R>                                                 \
+    struct op_impl<op_##id, op_l, B, L, R> {                                                      \
+        static char const *name() { return "__" #id "__"; }                                       \
+        static auto execute(const L &l, const R &r) -> decltype(expr) { return (expr); }          \
+        static B execute_cast(const L &l, const R &r) { return B(expr); }                         \
+    };                                                                                            \
+    template <typename B, typename L, typename R>                                                 \
+    struct op_impl<op_##id, op_r, B, L, R> {                                                      \
+        static char const *name() { return "__" #rid "__"; }                                      \
+        static auto execute(const R &r, const L &l) -> decltype(expr) { return (expr); }          \
+        static B execute_cast(const R &r, const L &l) { return B(expr); }                         \
+    };                                                                                            \
+    inline op_<op_##id, op_l, self_t, self_t> op(const self_t &, const self_t &) {                \
+        return op_<op_##id, op_l, self_t, self_t>();                                              \
+    }                                                                                             \
+    template <typename T>                                                                         \
+    op_<op_##id, op_l, self_t, T> op(const self_t &, const T &) {                                 \
+        return op_<op_##id, op_l, self_t, T>();                                                   \
+    }                                                                                             \
+    template <typename T>                                                                         \
+    op_<op_##id, op_r, T, self_t> op(const T &, const self_t &) {                                 \
+        return op_<op_##id, op_r, T, self_t>();                                                   \
+    }
+
+#define PYBIND11_INPLACE_OPERATOR(id, op, expr)                                                   \
+    template <typename B, typename L, typename R>                                                 \
+    struct op_impl<op_##id, op_l, B, L, R> {                                                      \
+        static char const *name() { return "__" #id "__"; }                                       \
+        static auto execute(L &l, const R &r) -> decltype(expr) { return expr; }                  \
+        static B execute_cast(L &l, const R &r) { return B(expr); }                               \
+    };                                                                                            \
+    template <typename T>                                                                         \
+    op_<op_##id, op_l, self_t, T> op(const self_t &, const T &) {                                 \
+        return op_<op_##id, op_l, self_t, T>();                                                   \
+    }
+
+#define PYBIND11_UNARY_OPERATOR(id, op, expr)                                                     \
+    template <typename B, typename L>                                                             \
+    struct op_impl<op_##id, op_u, B, L, undefined_t> {                                            \
+        static char const *name() { return "__" #id "__"; }                                       \
+        static auto execute(const L &l) -> decltype(expr) { return expr; }                        \
+        static B execute_cast(const L &l) { return B(expr); }                                     \
+    };                                                                                            \
+    inline op_<op_##id, op_u, self_t, undefined_t> op(const self_t &) {                           \
+        return op_<op_##id, op_u, self_t, undefined_t>();                                         \
+    }
+
+PYBIND11_BINARY_OPERATOR(sub, rsub, operator-, l - r)
+PYBIND11_BINARY_OPERATOR(add, radd, operator+, l + r)
+PYBIND11_BINARY_OPERATOR(mul, rmul, operator*, l *r)
+PYBIND11_BINARY_OPERATOR(truediv, rtruediv, operator/, l / r)
+PYBIND11_BINARY_OPERATOR(mod, rmod, operator%, l % r)
+PYBIND11_BINARY_OPERATOR(lshift, rlshift, operator<<, l << r)
+PYBIND11_BINARY_OPERATOR(rshift, rrshift, operator>>, l >> r)
+PYBIND11_BINARY_OPERATOR(and, rand, operator&, l &r)
+PYBIND11_BINARY_OPERATOR(xor, rxor, operator^, l ^ r)
+PYBIND11_BINARY_OPERATOR(eq, eq, operator==, l == r)
+PYBIND11_BINARY_OPERATOR(ne, ne, operator!=, l != r)
+PYBIND11_BINARY_OPERATOR(or, ror, operator|, l | r)
+PYBIND11_BINARY_OPERATOR(gt, lt, operator>, l > r)
+PYBIND11_BINARY_OPERATOR(ge, le, operator>=, l >= r)
+PYBIND11_BINARY_OPERATOR(lt, gt, operator<, l < r)
+PYBIND11_BINARY_OPERATOR(le, ge, operator<=, l <= r)
+// PYBIND11_BINARY_OPERATOR(pow,       rpow,         pow,          std::pow(l,  r))
+PYBIND11_INPLACE_OPERATOR(iadd, operator+=, l += r)
+PYBIND11_INPLACE_OPERATOR(isub, operator-=, l -= r)
+PYBIND11_INPLACE_OPERATOR(imul, operator*=, l *= r)
+PYBIND11_INPLACE_OPERATOR(itruediv, operator/=, l /= r)
+PYBIND11_INPLACE_OPERATOR(imod, operator%=, l %= r)
+PYBIND11_INPLACE_OPERATOR(ilshift, operator<<=, l <<= r)
+PYBIND11_INPLACE_OPERATOR(irshift, operator>>=, l >>= r)
+PYBIND11_INPLACE_OPERATOR(iand, operator&=, l &= r)
+PYBIND11_INPLACE_OPERATOR(ixor, operator^=, l ^= r)
+PYBIND11_INPLACE_OPERATOR(ior, operator|=, l |= r)
+PYBIND11_UNARY_OPERATOR(neg, operator-, -l)
+PYBIND11_UNARY_OPERATOR(pos, operator+, +l)
+// WARNING: This usage of `abs` should only be done for existing STL overloads.
+// Adding overloads directly in to the `std::` namespace is advised against:
+// https://en.cppreference.com/w/cpp/language/extending_std
+PYBIND11_UNARY_OPERATOR(abs, abs, std::abs(l))
+PYBIND11_UNARY_OPERATOR(hash, hash, std::hash<L>()(l))
+PYBIND11_UNARY_OPERATOR(invert, operator~, (~l))
+PYBIND11_UNARY_OPERATOR(bool, operator!, !!l)
+PYBIND11_UNARY_OPERATOR(int, int_, (int) l)
+PYBIND11_UNARY_OPERATOR(float, float_, (double) l)
+
+#undef PYBIND11_BINARY_OPERATOR
+#undef PYBIND11_INPLACE_OPERATOR
+#undef PYBIND11_UNARY_OPERATOR
+PYBIND11_NAMESPACE_END(detail)
+
+using detail::self;
+// Add named operators so that they are accessible via `py::`.
+using detail::hash;
+
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/options.h

@@ -0,0 +1,92 @@
+/*
+    pybind11/options.h: global settings that are configurable at runtime.
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "detail/common.h"
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+
+class options {
+public:
+    // Default RAII constructor, which leaves settings as they currently are.
+    options() : previous_state(global_state()) {}
+
+    // Class is non-copyable.
+    options(const options &) = delete;
+    options &operator=(const options &) = delete;
+
+    // Destructor, which restores settings that were in effect before.
+    ~options() { global_state() = previous_state; }
+
+    // Setter methods (affect the global state):
+
+    options &disable_user_defined_docstrings() & {
+        global_state().show_user_defined_docstrings = false;
+        return *this;
+    }
+
+    options &enable_user_defined_docstrings() & {
+        global_state().show_user_defined_docstrings = true;
+        return *this;
+    }
+
+    options &disable_function_signatures() & {
+        global_state().show_function_signatures = false;
+        return *this;
+    }
+
+    options &enable_function_signatures() & {
+        global_state().show_function_signatures = true;
+        return *this;
+    }
+
+    options &disable_enum_members_docstring() & {
+        global_state().show_enum_members_docstring = false;
+        return *this;
+    }
+
+    options &enable_enum_members_docstring() & {
+        global_state().show_enum_members_docstring = true;
+        return *this;
+    }
+
+    // Getter methods (return the global state):
+
+    static bool show_user_defined_docstrings() {
+        return global_state().show_user_defined_docstrings;
+    }
+
+    static bool show_function_signatures() { return global_state().show_function_signatures; }
+
+    static bool show_enum_members_docstring() {
+        return global_state().show_enum_members_docstring;
+    }
+
+    // This type is not meant to be allocated on the heap.
+    void *operator new(size_t) = delete;
+
+private:
+    struct state {
+        bool show_user_defined_docstrings = true; //< Include user-supplied texts in docstrings.
+        bool show_function_signatures = true;     //< Include auto-generated function signatures
+                                                  //  in docstrings.
+        bool show_enum_members_docstring = true;  //< Include auto-generated member list in enum
+                                                  //  docstrings.
+    };
+
+    static state &global_state() {
+        static state instance;
+        return instance;
+    }
+
+    state previous_state;
+};
+
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)

phivenv/Lib/site-packages/torch/include/pybind11/stl.h

@@ -0,0 +1,448 @@
+/*
+    pybind11/stl.h: Transparent conversion for STL data types
+
+    Copyright (c) 2016 Wenzel Jakob <wenzel.jakob@epfl.ch>
+
+    All rights reserved. Use of this source code is governed by a
+    BSD-style license that can be found in the LICENSE file.
+*/
+
+#pragma once
+
+#include "pybind11.h"
+#include "detail/common.h"
+
+#include <deque>
+#include <list>
+#include <map>
+#include <ostream>
+#include <set>
+#include <unordered_map>
+#include <unordered_set>
+#include <valarray>
+
+// See `detail/common.h` for implementation of these guards.
+#if defined(PYBIND11_HAS_OPTIONAL)
+#    include <optional>
+#elif defined(PYBIND11_HAS_EXP_OPTIONAL)
+#    include <experimental/optional>
+#endif
+
+#if defined(PYBIND11_HAS_VARIANT)
+#    include <variant>
+#endif
+
+PYBIND11_NAMESPACE_BEGIN(PYBIND11_NAMESPACE)
+PYBIND11_NAMESPACE_BEGIN(detail)
+
+/// Extracts an const lvalue reference or rvalue reference for U based on the type of T (e.g. for
+/// forwarding a container element).  Typically used indirect via forwarded_type(), below.
+template <typename T, typename U>
+using forwarded_type = conditional_t<std::is_lvalue_reference<T>::value,
+                                     remove_reference_t<U> &,
+                                     remove_reference_t<U> &&>;
+
+/// Forwards a value U as rvalue or lvalue according to whether T is rvalue or lvalue; typically
+/// used for forwarding a container's elements.
+template <typename T, typename U>
+constexpr forwarded_type<T, U> forward_like(U &&u) {
+    return std::forward<detail::forwarded_type<T, U>>(std::forward<U>(u));
+}
+
+// Checks if a container has a STL style reserve method.
+// This will only return true for a `reserve()` with a `void` return.
+template <typename C>
+using has_reserve_method = std::is_same<decltype(std::declval<C>().reserve(0)), void>;
+
+template <typename Type, typename Key>
+struct set_caster {
+    using type = Type;
+    using key_conv = make_caster<Key>;
+
+private:
+    template <typename T = Type, enable_if_t<has_reserve_method<T>::value, int> = 0>
+    void reserve_maybe(const anyset &s, Type *) {
+        value.reserve(s.size());
+    }
+    void reserve_maybe(const anyset &, void *) {}
+
+public:
+    bool load(handle src, bool convert) {
+        if (!isinstance<anyset>(src)) {
+            return false;
+        }
+        auto s = reinterpret_borrow<anyset>(src);
+        value.clear();
+        reserve_maybe(s, &value);
+        for (auto entry : s) {
+            key_conv conv;
+            if (!conv.load(entry, convert)) {
+                return false;
+            }
+            value.insert(cast_op<Key &&>(std::move(conv)));
+        }
+        return true;
+    }
+
+    template <typename T>
+    static handle cast(T &&src, return_value_policy policy, handle parent) {
+        if (!std::is_lvalue_reference<T>::value) {
+            policy = return_value_policy_override<Key>::policy(policy);
+        }
+        pybind11::set s;
+        for (auto &&value : src) {
+            auto value_ = reinterpret_steal<object>(
+                key_conv::cast(detail::forward_like<T>(value), policy, parent));
+            if (!value_ || !s.add(std::move(value_))) {
+                return handle();
+            }
+        }
+        return s.release();
+    }
+
+    PYBIND11_TYPE_CASTER(type, const_name("set[") + key_conv::name + const_name("]"));
+};
+
+template <typename Type, typename Key, typename Value>
+struct map_caster {
+    using key_conv = make_caster<Key>;
+    using value_conv = make_caster<Value>;
+
+private:
+    template <typename T = Type, enable_if_t<has_reserve_method<T>::value, int> = 0>
+    void reserve_maybe(const dict &d, Type *) {
+        value.reserve(d.size());
+    }
+    void reserve_maybe(const dict &, void *) {}
+
+public:
+    bool load(handle src, bool convert) {
+        if (!isinstance<dict>(src)) {
+            return false;
+        }
+        auto d = reinterpret_borrow<dict>(src);
+        value.clear();
+        reserve_maybe(d, &value);
+        for (auto it : d) {
+            key_conv kconv;
+            value_conv vconv;
+            if (!kconv.load(it.first.ptr(), convert) || !vconv.load(it.second.ptr(), convert)) {
+                return false;
+            }
+            value.emplace(cast_op<Key &&>(std::move(kconv)), cast_op<Value &&>(std::move(vconv)));
+        }
+        return true;
+    }
+
+    template <typename T>
+    static handle cast(T &&src, return_value_policy policy, handle parent) {
+        dict d;
+        return_value_policy policy_key = policy;
+        return_value_policy policy_value = policy;
+        if (!std::is_lvalue_reference<T>::value) {
+            policy_key = return_value_policy_override<Key>::policy(policy_key);
+            policy_value = return_value_policy_override<Value>::policy(policy_value);
+        }
+        for (auto &&kv : src) {
+            auto key = reinterpret_steal<object>(
+                key_conv::cast(detail::forward_like<T>(kv.first), policy_key, parent));
+            auto value = reinterpret_steal<object>(
+                value_conv::cast(detail::forward_like<T>(kv.second), policy_value, parent));
+            if (!key || !value) {
+                return handle();
+            }
+            d[std::move(key)] = std::move(value);
+        }
+        return d.release();
+    }
+
+    PYBIND11_TYPE_CASTER(Type,
+                         const_name("dict[") + key_conv::name + const_name(", ") + value_conv::name
+                             + const_name("]"));
+};
+
+template <typename Type, typename Value>
+struct list_caster {
+    using value_conv = make_caster<Value>;
+
+    bool load(handle src, bool convert) {
+        if (!isinstance<sequence>(src) || isinstance<bytes>(src) || isinstance<str>(src)) {
+            return false;
+        }
+        auto s = reinterpret_borrow<sequence>(src);
+        value.clear();
+        reserve_maybe(s, &value);
+        for (const auto &it : s) {
+            value_conv conv;
+            if (!conv.load(it, convert)) {
+                return false;
+            }
+            value.push_back(cast_op<Value &&>(std::move(conv)));
+        }
+        return true;
+    }
+
+private:
+    template <typename T = Type, enable_if_t<has_reserve_method<T>::value, int> = 0>
+    void reserve_maybe(const sequence &s, Type *) {
+        value.reserve(s.size());
+    }
+    void reserve_maybe(const sequence &, void *) {}
+
+public:
+    template <typename T>
+    static handle cast(T &&src, return_value_policy policy, handle parent) {
+        if (!std::is_lvalue_reference<T>::value) {
+            policy = return_value_policy_override<Value>::policy(policy);
+        }
+        list l(src.size());
+        ssize_t index = 0;
+        for (auto &&value : src) {
+            auto value_ = reinterpret_steal<object>(
+                value_conv::cast(detail::forward_like<T>(value), policy, parent));
+            if (!value_) {
+                return handle();
+            }
+            PyList_SET_ITEM(l.ptr(), index++, value_.release().ptr()); // steals a reference
+        }
+        return l.release();
+    }
+
+    PYBIND11_TYPE_CASTER(Type, const_name("list[") + value_conv::name + const_name("]"));
+};
+
+template <typename Type, typename Alloc>
+struct type_caster<std::vector<Type, Alloc>> : list_caster<std::vector<Type, Alloc>, Type> {};
+
+template <typename Type, typename Alloc>
+struct type_caster<std::deque<Type, Alloc>> : list_caster<std::deque<Type, Alloc>, Type> {};
+
+template <typename Type, typename Alloc>
+struct type_caster<std::list<Type, Alloc>> : list_caster<std::list<Type, Alloc>, Type> {};
+
+template <typename ArrayType, typename Value, bool Resizable, size_t Size = 0>
+struct array_caster {
+    using value_conv = make_caster<Value>;
+
+private:
+    template <bool R = Resizable>
+    bool require_size(enable_if_t<R, size_t> size) {
+        if (value.size() != size) {
+            value.resize(size);
+        }
+        return true;
+    }
+    template <bool R = Resizable>
+    bool require_size(enable_if_t<!R, size_t> size) {
+        return size == Size;
+    }
+
+public:
+    bool load(handle src, bool convert) {
+        if (!isinstance<sequence>(src)) {
+            return false;
+        }
+        auto l = reinterpret_borrow<sequence>(src);
+        if (!require_size(l.size())) {
+            return false;
+        }
+        size_t ctr = 0;
+        for (const auto &it : l) {
+            value_conv conv;
+            if (!conv.load(it, convert)) {
+                return false;
+            }
+            value[ctr++] = cast_op<Value &&>(std::move(conv));
+        }
+        return true;
+    }
+
+    template <typename T>
+    static handle cast(T &&src, return_value_policy policy, handle parent) {
+        list l(src.size());
+        ssize_t index = 0;
+        for (auto &&value : src) {
+            auto value_ = reinterpret_steal<object>(
+                value_conv::cast(detail::forward_like<T>(value), policy, parent));
+            if (!value_) {
+                return handle();
+            }
+            PyList_SET_ITEM(l.ptr(), index++, value_.release().ptr()); // steals a reference
+        }
+        return l.release();
+    }
+
+    PYBIND11_TYPE_CASTER(ArrayType,
+                         const_name<Resizable>(const_name(""), const_name("Annotated["))
+                             + const_name("list[") + value_conv::name + const_name("]")
+                             + const_name<Resizable>(const_name(""),
+                                                     const_name(", FixedSize(")
+                                                         + const_name<Size>() + const_name(")]")));
+};
+
+template <typename Type, size_t Size>
+struct type_caster<std::array<Type, Size>>
+    : array_caster<std::array<Type, Size>, Type, false, Size> {};
+
+template <typename Type>
+struct type_caster<std::valarray<Type>> : array_caster<std::valarray<Type>, Type, true> {};
+
+template <typename Key, typename Compare, typename Alloc>
+struct type_caster<std::set<Key, Compare, Alloc>>
+    : set_caster<std::set<Key, Compare, Alloc>, Key> {};
+
+template <typename Key, typename Hash, typename Equal, typename Alloc>
+struct type_caster<std::unordered_set<Key, Hash, Equal, Alloc>>
+    : set_caster<std::unordered_set<Key, Hash, Equal, Alloc>, Key> {};
+
+template <typename Key, typename Value, typename Compare, typename Alloc>
+struct type_caster<std::map<Key, Value, Compare, Alloc>>
+    : map_caster<std::map<Key, Value, Compare, Alloc>, Key, Value> {};
+
+template <typename Key, typename Value, typename Hash, typename Equal, typename Alloc>
+struct type_caster<std::unordered_map<Key, Value, Hash, Equal, Alloc>>
+    : map_caster<std::unordered_map<Key, Value, Hash, Equal, Alloc>, Key, Value> {};
+
+// This type caster is intended to be used for std::optional and std::experimental::optional
+template <typename Type, typename Value = typename Type::value_type>
+struct optional_caster {
+    using value_conv = make_caster<Value>;
+
+    template <typename T>
+    static handle cast(T &&src, return_value_policy policy, handle parent) {
+        if (!src) {
+            return none().release();
+        }
+        if (!std::is_lvalue_reference<T>::value) {
+            policy = return_value_policy_override<Value>::policy(policy);
+        }
+        // NOLINTNEXTLINE(bugprone-unchecked-optional-access)
+        return value_conv::cast(*std::forward<T>(src), policy, parent);
+    }
+
+    bool load(handle src, bool convert) {
+        if (!src) {
+            return false;
+        }
+        if (src.is_none()) {
+            return true; // default-constructed value is already empty
+        }
+        value_conv inner_caster;
+        if (!inner_caster.load(src, convert)) {
+            return false;
+        }
+
+        value.emplace(cast_op<Value &&>(std::move(inner_caster)));
+        return true;
+    }
+
+    PYBIND11_TYPE_CASTER(Type, const_name("Optional[") + value_conv::name + const_name("]"));
+};
+
+#if defined(PYBIND11_HAS_OPTIONAL)
+template <typename T>
+struct type_caster<std::optional<T>> : public optional_caster<std::optional<T>> {};
+
+template <>
+struct type_caster<std::nullopt_t> : public void_caster<std::nullopt_t> {};
+#endif
+
+#if defined(PYBIND11_HAS_EXP_OPTIONAL)
+template <typename T>
+struct type_caster<std::experimental::optional<T>>
+    : public optional_caster<std::experimental::optional<T>> {};
+
+template <>
+struct type_caster<std::experimental::nullopt_t>
+    : public void_caster<std::experimental::nullopt_t> {};
+#endif
+
+/// Visit a variant and cast any found type to Python
+struct variant_caster_visitor {
+    return_value_policy policy;
+    handle parent;
+
+    using result_type = handle; // required by boost::variant in C++11
+
+    template <typename T>
+    result_type operator()(T &&src) const {
+        return make_caster<T>::cast(std::forward<T>(src), policy, parent);
+    }
+};
+
+/// Helper class which abstracts away variant's `visit` function. `std::variant` and similar
+/// `namespace::variant` types which provide a `namespace::visit()` function are handled here
+/// automatically using argument-dependent lookup. Users can provide specializations for other
+/// variant-like classes, e.g. `boost::variant` and `boost::apply_visitor`.
+template <template <typename...> class Variant>
+struct visit_helper {
+    template <typename... Args>
+    static auto call(Args &&...args) -> decltype(visit(std::forward<Args>(args)...)) {
+        return visit(std::forward<Args>(args)...);
+    }
+};
+
+/// Generic variant caster
+template <typename Variant>
+struct variant_caster;
+
+template <template <typename...> class V, typename... Ts>
+struct variant_caster<V<Ts...>> {
+    static_assert(sizeof...(Ts) > 0, "Variant must consist of at least one alternative.");
+
+    template <typename U, typename... Us>
+    bool load_alternative(handle src, bool convert, type_list<U, Us...>) {
+        auto caster = make_caster<U>();
+        if (caster.load(src, convert)) {
+            value = cast_op<U>(std::move(caster));
+            return true;
+        }
+        return load_alternative(src, convert, type_list<Us...>{});
+    }
+
+    bool load_alternative(handle, bool, type_list<>) { return false; }
+
+    bool load(handle src, bool convert) {
+        // Do a first pass without conversions to improve constructor resolution.
+        // E.g. `py::int_(1).cast<variant<double, int>>()` needs to fill the `int`
+        // slot of the variant. Without two-pass loading `double` would be filled
+        // because it appears first and a conversion is possible.
+        if (convert && load_alternative(src, false, type_list<Ts...>{})) {
+            return true;
+        }
+        return load_alternative(src, convert, type_list<Ts...>{});
+    }
+
+    template <typename Variant>
+    static handle cast(Variant &&src, return_value_policy policy, handle parent) {
+        return visit_helper<V>::call(variant_caster_visitor{policy, parent},
+                                     std::forward<Variant>(src));
+    }
+
+    using Type = V<Ts...>;
+    PYBIND11_TYPE_CASTER(Type,
+                         const_name("Union[")
+                             + ::pybind11::detail::concat(make_caster<Ts>::name...)
+                             + const_name("]"));
+};
+
+#if defined(PYBIND11_HAS_VARIANT)
+template <typename... Ts>
+struct type_caster<std::variant<Ts...>> : variant_caster<std::variant<Ts...>> {};
+
+template <>
+struct type_caster<std::monostate> : public void_caster<std::monostate> {};
+#endif
+
+PYBIND11_NAMESPACE_END(detail)
+
+inline std::ostream &operator<<(std::ostream &os, const handle &obj) {
+#ifdef PYBIND11_HAS_STRING_VIEW
+    os << str(obj).cast<std::string_view>();
+#else
+    os << (std::string) str(obj);
+#endif
+    return os;
+}
+
+PYBIND11_NAMESPACE_END(PYBIND11_NAMESPACE)