Add files using upload-large-folder tool

.gitattributes

@@ -75,3 +75,4 @@ tuning-competition-baseline/.venv/lib/python3.11/site-packages/pip/_vendor/pkg_r
 tuning-competition-baseline/.venv/lib/python3.11/site-packages/nvidia/cufft/lib/libcufftw.so.10 filter=lfs diff=lfs merge=lfs -text
 tuning-competition-baseline/.venv/lib/python3.11/site-packages/Cython/Compiler/__pycache__/ModuleNode.cpython-311.pyc filter=lfs diff=lfs merge=lfs -text
 tuning-competition-baseline/.venv/lib/python3.11/site-packages/nvidia/cublas/lib/libnvblas.so.11 filter=lfs diff=lfs merge=lfs -text
+tuning-competition-baseline/.venv/lib/python3.11/site-packages/pip/_vendor/distlib/t64.exe filter=lfs diff=lfs merge=lfs -text

tuning-competition-baseline/.venv/lib/python3.11/site-packages/pip/_vendor/distlib/t64.exe

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81a618f21cb87db9076134e70388b6e9cb7c2106739011b6a51772d22cae06b7
+size 108032

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/_inductor/freezing.py

@@ -0,0 +1,266 @@
+from __future__ import annotations
+
+import itertools
+import logging
+
+import weakref
+from typing import Any, List, Optional, Tuple
+
+import torch
+import torch.utils._pytree as pytree
+from torch._dynamo.utils import dynamo_timed, lazy_format_graph_code
+from torch._functorch.aot_autograd import MutationType
+from torch._functorch.compile_utils import fx_graph_cse
+from torch._inductor.constant_folding import constant_fold, replace_node_with_constant
+
+from torch._inductor.fx_passes.freezing_patterns import freezing_passes
+from torch._inductor.fx_passes.post_grad import view_to_reshape
+
+from . import config
+
+aten = torch.ops.aten
+prims = torch.ops.prims
+
+log = logging.getLogger(__name__)
+
+
+def replace_params_with_constants(
+    gm: torch.fx.GraphModule,
+    flat_params: list[Any],
+    fw_metadata: torch._functorch.aot_autograd.ViewAndMutationMeta,
+) -> List[int]:
+    """
+    Replaces the parameters of a PyTorch GraphModule with constants wherever possible.
+    Returns a list of indices representing the input parameters that were not converted to constants.
+    """
+    params = [node for node in gm.graph.nodes if node.op == "placeholder"]
+    fake_inp_nodes = params[: len(params)]
+    preserved_arg_indices = []
+    aliased_input_args = [
+        out_info.base_idx
+        for out_info in fw_metadata.output_info
+        if out_info.base_idx is not None
+    ]
+
+    # TODO (tmanlaibaatar) figure out why this is different
+    # from mutated_inp_runtime_indices
+    mutated_inps = [
+        i
+        for i, m in enumerate(fw_metadata.input_info)
+        if m.mutation_type
+        in (MutationType.MUTATED_IN_GRAPH, MutationType.MUTATED_OUT_GRAPH)
+    ]
+
+    for i, (real_input, node) in enumerate(zip(flat_params, fake_inp_nodes)):
+        if i in mutated_inps or i in aliased_input_args:
+            preserved_arg_indices.append(i)
+            continue
+        replace_node_with_constant(gm, node, real_input)
+    # add on non param inputs
+    preserved_arg_indices.extend(range(len(flat_params), len(params)))
+    # is this necessary ?
+    gm.recompile()
+    return preserved_arg_indices
+
+
+def freeze(
+    dynamo_gm: torch.fx.GraphModule,
+    aot_autograd_gm: torch.fx.GraphModule,
+    example_inputs: List[torch._subclasses.FakeTensor],
+) -> Tuple[torch.fx.GraphModule, List[int]]:
+    """
+    Inlines parameters that are not mutated into constants and optimizes the graph through constant propagation
+    and other techniques. If enabled, the function also discards the original parameters of the module for memory efficiency.
+
+    Assumes that this function is run in dynamo tracing post aot_autograd.
+
+    Args:
+        dynamo_gm (torch.fx.GraphModule): The Dynamo constructed GraphModule.
+        aot_autograd_gm (torch.fx.GraphModule): The aot_autograd constructed GraphModule to be frozen.
+        example_inputs (List[torch.Tensor]): A list of example input tensors to be used in the freezing process.
+
+    Returns:
+        Tuple[torch.fx.GraphModule, List[int]]: A tuple containing the frozen GraphModule and a list of indices
+        of the inputs that were preserved (not turned into constants).
+    """
+    # We have convert conv's weight to channels last which may meet error for .view
+    # when doing fake_tensor_prop. So we need to convert view to reshape first.
+    # See the details in fx_codegen_and_compile of compile_fx.py.
+    view_to_reshape(aot_autograd_gm)
+
+    if tracing_context := torch._guards.TracingContext.try_get():
+        fw_metadata = tracing_context.fw_metadata
+        params_flat = tracing_context.params_flat
+        assert fw_metadata is not None and params_flat is not None
+
+        preserved_arg_indices = replace_params_with_constants(
+            aot_autograd_gm, params_flat, fw_metadata
+        )
+    else:
+        inputs = [
+            node for node in aot_autograd_gm.graph.nodes if node.op == "placeholder"
+        ]
+        preserved_arg_indices = list(range(len(inputs)))
+
+    # TODO - further restrict cse ? right now needed to dedup aliasing ops
+    cse_graph = fx_graph_cse(aot_autograd_gm.graph)
+    aot_autograd_gm.graph = cse_graph
+    aot_autograd_gm.recompile()
+
+    aot_example_inputs = [example_inputs[ind] for ind in preserved_arg_indices]
+    freezing_passes(aot_autograd_gm, aot_example_inputs)
+
+    constant_fold(aot_autograd_gm)
+    # invalidate nn Modules
+    if config.freezing_discard_parameters:
+        invalidate_eager_modules()
+        discard_traced_gm_params(dynamo_gm)
+
+    log.debug("%s", lazy_format_graph_code("FROZEN GRAPH", aot_autograd_gm))
+
+    return aot_autograd_gm, preserved_arg_indices
+
+
+class ErasedTensor(torch.Tensor):
+    @staticmethod
+    def __new__(cls, elem, name, owning_mod):
+        return super().__new__(cls, elem.to(device="meta"))
+
+    def __init__(self, elem, name: Optional[str], mod):
+        self.erased_name = name
+        self.owning_mod_ref = weakref.ref(mod)
+
+    @classmethod
+    def __torch_dispatch__(cls, func, types, args=(), kwargs=None):
+        erased_tensors = [
+            e
+            for e in pytree.arg_tree_leaves(*args, **kwargs)
+            if isinstance(e, ErasedTensor)
+        ]
+        assert len(erased_tensors) > 0
+        e = erased_tensors[0]
+
+        raise RuntimeError(
+            f"Trying to run Pytorch Eager Module after Dynamo Freezing. "
+            "The original parameters have been discarded for memory efficiency. "
+            f"Found in op {func} for erased parameter {e.erased_name} of {e.owning_mod_ref()}"
+        )
+
+
+@torch.utils._python_dispatch._disable_current_modes()
+def invalidate_eager_modules():
+    for mod in torch._guards.TracingContext.get().module_context.nn_modules.values():
+        if not isinstance(mod, torch.nn.Module):
+            continue
+
+        for attr_name, tensor in list(
+            itertools.chain(
+                mod.named_parameters(recurse=False), mod.named_buffers(recurse=False)
+            )
+        ):
+            with torch._dispatch.python.no_python_dispatcher():
+                e_t = ErasedTensor(tensor, attr_name, mod)
+            if isinstance(tensor, torch.nn.Parameter):
+                e_t.requires_grad_(True)
+                e_t._is_param = True  # type: ignore[attr-defined]
+            setattr(mod, attr_name, e_t)
+
+
+@torch.utils._python_dispatch._disable_current_modes()
+def discard_traced_gm_params(mod: torch.fx.GraphModule):
+    for attr_name, tensor in list(
+        itertools.chain(
+            mod.named_parameters(recurse=False), mod.named_buffers(recurse=False)
+        )
+    ):
+        with torch._dispatch.python.no_python_dispatcher():
+            e_t = ErasedTensor(tensor, attr_name, mod)
+        if isinstance(tensor, torch.nn.Parameter):
+            e_t.requires_grad_(True)
+            e_t._is_param = True  # type: ignore[attr-defined]
+        setattr(mod, attr_name, e_t)
+
+
+def enforce_output_layout(gm: torch.fx.GraphModule):
+    """
+    Make sure the output node's layout does not change due to compiler optimizations
+    by adding aten.as_strided nodes with the expected strides.
+
+    Only used for inference so we can assume all graph outputs are model outputs.
+    """
+    *_, output_node = gm.graph.nodes
+    out_list = output_node.args[0]
+    with gm.graph.inserting_before(output_node):
+        for n in out_list:
+            if not isinstance(
+                n.meta["val"], torch.Tensor
+            ) or not torch._prims_common.is_non_overlapping_and_dense(n.meta["val"]):
+                continue
+
+            # add a node to enforce eager layout
+            ft = n.meta["val"]
+            new_node = gm.graph.call_function(
+                prims.inductor_force_stride_order.default, (n, ft.stride())
+            )
+
+            # can not call
+            # n.replace_all_uses_with(new_node)
+            # since it will replace the usage of n in new_node itself.
+            output_node.replace_input_with(n, new_node)
+
+    gm.graph.lint()
+    gm.recompile()
+
+
+def enforce_as_strided_input_layout(gm: torch.fx.GraphModule):
+    """
+    Make sure the as_strided node's input's layout does not change due to compiler
+    optimizations, because the as_strided strides info depends on input tensor stride info.
+    """
+
+    as_strided_ops = [
+        torch.ops.aten.as_strided.default,
+        torch.ops.aten.as_strided_.default,
+        torch.ops.aten.as_strided_scatter.default,
+    ]
+    strided_nodes = [n for n in gm.graph.nodes if n.target in as_strided_ops]
+    for n in strided_nodes:
+        with gm.graph.inserting_before(n):
+            # add a node to enforce eager layout
+            ft = n.args[0].meta["val"]
+            new_node = gm.graph.call_function(
+                prims.inductor_force_stride_order.default, (n.args[0], ft.stride())
+            )
+            n.replace_input_with(n.args[0], new_node)
+
+    gm.graph.lint()
+    gm.recompile()
+
+
+@dynamo_timed
+def convert_conv_weights_to_channels_last(gm: torch.fx.GraphModule):
+    """
+    Convert 4d convolution weight tensor to channels last format.
+
+    This pass is performed before freezing so the added nodes can be constant
+    folded by freezing.
+    """
+    convs = [n for n in gm.graph.nodes if n.target == aten.convolution.default]
+    for conv in convs:
+        weight_node = conv.args[1]
+        if len(weight_node.meta["val"].size()) != 4 or weight_node.meta[
+            "val"
+        ].is_contiguous(memory_format=torch.channels_last):
+            # not a 4d tensor or already channels last, skip
+            continue
+
+        with gm.graph.inserting_before(conv):
+            new_node = gm.graph.call_function(
+                aten.clone.default,
+                (weight_node,),
+                {"memory_format": torch.channels_last},
+            )
+            conv.replace_input_with(weight_node, new_node)
+
+    enforce_as_strided_input_layout(gm)
+    enforce_output_layout(gm)

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/_inductor/quantized_lowerings.py

@@ -0,0 +1,15 @@
+import torch
+
+
+def register_quantized_ops():
+    from . import lowering
+
+    quantized = torch.ops.quantized
+
+    lowering.add_needs_realized_inputs(
+        [
+            quantized.max_pool2d,
+        ]
+    )
+
+    lowering.make_fallback(quantized.max_pool2d)

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/Backtrace.h

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

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/CPUFunctions_inl.h

@@ -0,0 +1,576 @@
+#pragma once
+// @generated by torchgen/gen.py from DispatchKeyFunctions_inl.h
+
+// NB: The implementing C++ file is RegisterDispatchKey.cpp
+
+// The only #includes we need are for custom classes that have defaults in the C++ API
+#include <c10/core/MemoryFormat.h>
+#include <c10/core/Scalar.h>
+#include <ATen/core/Reduction.h>
+
+#if defined(AT_PER_OPERATOR_HEADERS) && defined(TORCH_ASSERT_ONLY_METHOD_OPERATORS)
+#error This change adds a dependency on all pytorch operators, meaning the     \
+  file will need to be re-compiled every time an operator is changed or added. \
+  Consider including a specific operator from                                  \
+  <ATen/ops/{my_operator}_cpu_dispatch.h>.                   \
+  See NOTE [TORCH_ASSERT_ONLY_METHOD_OPERATORS].
+#endif
+
+#include <ATen/ops/_adaptive_avg_pool2d_cpu_dispatch.h>
+#include <ATen/ops/_adaptive_avg_pool2d_backward_cpu_dispatch.h>
+#include <ATen/ops/_adaptive_avg_pool3d_cpu_dispatch.h>
+#include <ATen/ops/_adaptive_avg_pool3d_backward_cpu_dispatch.h>
+#include <ATen/ops/_add_relu_cpu_dispatch.h>
+#include <ATen/ops/_addmm_activation_cpu_dispatch.h>
+#include <ATen/ops/_aminmax_cpu_dispatch.h>
+#include <ATen/ops/_amp_foreach_non_finite_check_and_unscale_cpu_dispatch.h>
+#include <ATen/ops/_amp_update_scale_cpu_dispatch.h>
+#include <ATen/ops/_assert_async_cpu_dispatch.h>
+#include <ATen/ops/_cdist_backward_cpu_dispatch.h>
+#include <ATen/ops/_cdist_forward_cpu_dispatch.h>
+#include <ATen/ops/_cholesky_solve_helper_cpu_dispatch.h>
+#include <ATen/ops/_compute_linear_combination_cpu_dispatch.h>
+#include <ATen/ops/_convert_indices_from_coo_to_csr_cpu_dispatch.h>
+#include <ATen/ops/_convert_indices_from_csr_to_coo_cpu_dispatch.h>
+#include <ATen/ops/_convert_weight_to_int4pack_cpu_dispatch.h>
+#include <ATen/ops/_ctc_loss_cpu_dispatch.h>
+#include <ATen/ops/_ctc_loss_backward_cpu_dispatch.h>
+#include <ATen/ops/_cummax_helper_cpu_dispatch.h>
+#include <ATen/ops/_cummin_helper_cpu_dispatch.h>
+#include <ATen/ops/_dirichlet_grad_cpu_dispatch.h>
+#include <ATen/ops/_efficientzerotensor_cpu_dispatch.h>
+#include <ATen/ops/_embedding_bag_cpu_dispatch.h>
+#include <ATen/ops/_embedding_bag_dense_backward_cpu_dispatch.h>
+#include <ATen/ops/_embedding_bag_forward_only_cpu_dispatch.h>
+#include <ATen/ops/_embedding_bag_per_sample_weights_backward_cpu_dispatch.h>
+#include <ATen/ops/_empty_affine_quantized_cpu_dispatch.h>
+#include <ATen/ops/_empty_per_channel_affine_quantized_cpu_dispatch.h>
+#include <ATen/ops/_fake_quantize_learnable_per_channel_affine_cpu_dispatch.h>
+#include <ATen/ops/_fake_quantize_learnable_per_channel_affine_backward_cpu_dispatch.h>
+#include <ATen/ops/_fake_quantize_learnable_per_tensor_affine_cpu_dispatch.h>
+#include <ATen/ops/_fake_quantize_learnable_per_tensor_affine_backward_cpu_dispatch.h>
+#include <ATen/ops/_fake_quantize_per_tensor_affine_cachemask_tensor_qparams_cpu_dispatch.h>
+#include <ATen/ops/_fft_c2c_cpu_dispatch.h>
+#include <ATen/ops/_fft_c2r_cpu_dispatch.h>
+#include <ATen/ops/_fft_r2c_cpu_dispatch.h>
+#include <ATen/ops/_foobar_cpu_dispatch.h>
+#include <ATen/ops/_foreach_abs_cpu_dispatch.h>
+#include <ATen/ops/_foreach_acos_cpu_dispatch.h>
+#include <ATen/ops/_foreach_add_cpu_dispatch.h>
+#include <ATen/ops/_foreach_addcdiv_cpu_dispatch.h>
+#include <ATen/ops/_foreach_addcmul_cpu_dispatch.h>
+#include <ATen/ops/_foreach_asin_cpu_dispatch.h>
+#include <ATen/ops/_foreach_atan_cpu_dispatch.h>
+#include <ATen/ops/_foreach_ceil_cpu_dispatch.h>
+#include <ATen/ops/_foreach_clamp_max_cpu_dispatch.h>
+#include <ATen/ops/_foreach_clamp_min_cpu_dispatch.h>
+#include <ATen/ops/_foreach_copy_cpu_dispatch.h>
+#include <ATen/ops/_foreach_cos_cpu_dispatch.h>
+#include <ATen/ops/_foreach_cosh_cpu_dispatch.h>
+#include <ATen/ops/_foreach_div_cpu_dispatch.h>
+#include <ATen/ops/_foreach_erf_cpu_dispatch.h>
+#include <ATen/ops/_foreach_erfc_cpu_dispatch.h>
+#include <ATen/ops/_foreach_exp_cpu_dispatch.h>
+#include <ATen/ops/_foreach_expm1_cpu_dispatch.h>
+#include <ATen/ops/_foreach_floor_cpu_dispatch.h>
+#include <ATen/ops/_foreach_frac_cpu_dispatch.h>
+#include <ATen/ops/_foreach_lerp_cpu_dispatch.h>
+#include <ATen/ops/_foreach_lgamma_cpu_dispatch.h>
+#include <ATen/ops/_foreach_log_cpu_dispatch.h>
+#include <ATen/ops/_foreach_log10_cpu_dispatch.h>
+#include <ATen/ops/_foreach_log1p_cpu_dispatch.h>
+#include <ATen/ops/_foreach_log2_cpu_dispatch.h>
+#include <ATen/ops/_foreach_maximum_cpu_dispatch.h>
+#include <ATen/ops/_foreach_minimum_cpu_dispatch.h>
+#include <ATen/ops/_foreach_mul_cpu_dispatch.h>
+#include <ATen/ops/_foreach_neg_cpu_dispatch.h>
+#include <ATen/ops/_foreach_norm_cpu_dispatch.h>
+#include <ATen/ops/_foreach_pow_cpu_dispatch.h>
+#include <ATen/ops/_foreach_reciprocal_cpu_dispatch.h>
+#include <ATen/ops/_foreach_round_cpu_dispatch.h>
+#include <ATen/ops/_foreach_sigmoid_cpu_dispatch.h>
+#include <ATen/ops/_foreach_sign_cpu_dispatch.h>
+#include <ATen/ops/_foreach_sin_cpu_dispatch.h>
+#include <ATen/ops/_foreach_sinh_cpu_dispatch.h>
+#include <ATen/ops/_foreach_sqrt_cpu_dispatch.h>
+#include <ATen/ops/_foreach_sub_cpu_dispatch.h>
+#include <ATen/ops/_foreach_tan_cpu_dispatch.h>
+#include <ATen/ops/_foreach_tanh_cpu_dispatch.h>
+#include <ATen/ops/_foreach_trunc_cpu_dispatch.h>
+#include <ATen/ops/_foreach_zero_cpu_dispatch.h>
+#include <ATen/ops/_functional_assert_async_cpu_dispatch.h>
+#include <ATen/ops/_fused_moving_avg_obs_fq_helper_cpu_dispatch.h>
+#include <ATen/ops/_fused_sdp_choice_cpu_dispatch.h>
+#include <ATen/ops/_histogramdd_bin_edges_cpu_dispatch.h>
+#include <ATen/ops/_histogramdd_from_bin_cts_cpu_dispatch.h>
+#include <ATen/ops/_histogramdd_from_bin_tensors_cpu_dispatch.h>
+#include <ATen/ops/_index_put_impl_cpu_dispatch.h>
+#include <ATen/ops/_linalg_det_cpu_dispatch.h>
+#include <ATen/ops/_linalg_eigh_cpu_dispatch.h>
+#include <ATen/ops/_linalg_eigvals_cpu_dispatch.h>
+#include <ATen/ops/_linalg_slogdet_cpu_dispatch.h>
+#include <ATen/ops/_linalg_solve_ex_cpu_dispatch.h>
+#include <ATen/ops/_linalg_svd_cpu_dispatch.h>
+#include <ATen/ops/_local_scalar_dense_cpu_dispatch.h>
+#include <ATen/ops/_log_softmax_cpu_dispatch.h>
+#include <ATen/ops/_log_softmax_backward_data_cpu_dispatch.h>
+#include <ATen/ops/_logcumsumexp_cpu_dispatch.h>
+#include <ATen/ops/_make_dep_token_cpu_dispatch.h>
+#include <ATen/ops/_make_per_channel_quantized_tensor_cpu_dispatch.h>
+#include <ATen/ops/_make_per_tensor_quantized_tensor_cpu_dispatch.h>
+#include <ATen/ops/_masked_softmax_cpu_dispatch.h>
+#include <ATen/ops/_masked_softmax_backward_cpu_dispatch.h>
+#include <ATen/ops/_native_batch_norm_legit_cpu_dispatch.h>
+#include <ATen/ops/_native_multi_head_attention_cpu_dispatch.h>
+#include <ATen/ops/_nested_from_padded_cpu_dispatch.h>
+#include <ATen/ops/_nested_tensor_from_mask_cpu_dispatch.h>
+#include <ATen/ops/_nested_tensor_from_mask_left_aligned_cpu_dispatch.h>
+#include <ATen/ops/_nested_view_from_buffer_cpu_dispatch.h>
+#include <ATen/ops/_pdist_backward_cpu_dispatch.h>
+#include <ATen/ops/_pdist_forward_cpu_dispatch.h>
+#include <ATen/ops/_prelu_kernel_cpu_dispatch.h>
+#include <ATen/ops/_prelu_kernel_backward_cpu_dispatch.h>
+#include <ATen/ops/_reshape_alias_cpu_dispatch.h>
+#include <ATen/ops/_sample_dirichlet_cpu_dispatch.h>
+#include <ATen/ops/_scaled_dot_product_flash_attention_for_cpu_cpu_dispatch.h>
+#include <ATen/ops/_scaled_dot_product_flash_attention_for_cpu_backward_cpu_dispatch.h>
+#include <ATen/ops/_segment_reduce_backward_cpu_dispatch.h>
+#include <ATen/ops/_slow_conv2d_backward_cpu_dispatch.h>
+#include <ATen/ops/_slow_conv2d_forward_cpu_dispatch.h>
+#include <ATen/ops/_softmax_cpu_dispatch.h>
+#include <ATen/ops/_softmax_backward_data_cpu_dispatch.h>
+#include <ATen/ops/_spdiags_cpu_dispatch.h>
+#include <ATen/ops/_stack_cpu_dispatch.h>
+#include <ATen/ops/_standard_gamma_cpu_dispatch.h>
+#include <ATen/ops/_standard_gamma_grad_cpu_dispatch.h>
+#include <ATen/ops/_test_functorch_fallback_cpu_dispatch.h>
+#include <ATen/ops/_test_optional_filled_intlist_cpu_dispatch.h>
+#include <ATen/ops/_test_optional_floatlist_cpu_dispatch.h>
+#include <ATen/ops/_test_optional_intlist_cpu_dispatch.h>
+#include <ATen/ops/_to_sparse_cpu_dispatch.h>
+#include <ATen/ops/_to_sparse_bsc_cpu_dispatch.h>
+#include <ATen/ops/_to_sparse_bsr_cpu_dispatch.h>
+#include <ATen/ops/_to_sparse_csc_cpu_dispatch.h>
+#include <ATen/ops/_to_sparse_csr_cpu_dispatch.h>
+#include <ATen/ops/_transform_bias_rescale_qkv_cpu_dispatch.h>
+#include <ATen/ops/_transformer_encoder_layer_fwd_cpu_dispatch.h>
+#include <ATen/ops/_unique_cpu_dispatch.h>
+#include <ATen/ops/_unique2_cpu_dispatch.h>
+#include <ATen/ops/_upsample_bicubic2d_aa_cpu_dispatch.h>
+#include <ATen/ops/_upsample_bicubic2d_aa_backward_cpu_dispatch.h>
+#include <ATen/ops/_upsample_bilinear2d_aa_cpu_dispatch.h>
+#include <ATen/ops/_upsample_bilinear2d_aa_backward_cpu_dispatch.h>
+#include <ATen/ops/_upsample_nearest_exact1d_cpu_dispatch.h>
+#include <ATen/ops/_upsample_nearest_exact1d_backward_cpu_dispatch.h>
+#include <ATen/ops/_upsample_nearest_exact2d_cpu_dispatch.h>
+#include <ATen/ops/_upsample_nearest_exact2d_backward_cpu_dispatch.h>
+#include <ATen/ops/_upsample_nearest_exact3d_cpu_dispatch.h>
+#include <ATen/ops/_upsample_nearest_exact3d_backward_cpu_dispatch.h>
+#include <ATen/ops/_validate_compressed_sparse_indices_cpu_dispatch.h>
+#include <ATen/ops/_weight_int4pack_mm_cpu_dispatch.h>
+#include <ATen/ops/_weight_int8pack_mm_cpu_dispatch.h>
+#include <ATen/ops/_weight_norm_interface_cpu_dispatch.h>
+#include <ATen/ops/_weight_norm_interface_backward_cpu_dispatch.h>
+#include <ATen/ops/abs_cpu_dispatch.h>
+#include <ATen/ops/acos_cpu_dispatch.h>
+#include <ATen/ops/acosh_cpu_dispatch.h>
+#include <ATen/ops/adaptive_avg_pool2d_cpu_dispatch.h>
+#include <ATen/ops/adaptive_avg_pool3d_cpu_dispatch.h>
+#include <ATen/ops/adaptive_avg_pool3d_backward_cpu_dispatch.h>
+#include <ATen/ops/adaptive_max_pool2d_cpu_dispatch.h>
+#include <ATen/ops/adaptive_max_pool2d_backward_cpu_dispatch.h>
+#include <ATen/ops/adaptive_max_pool3d_cpu_dispatch.h>
+#include <ATen/ops/adaptive_max_pool3d_backward_cpu_dispatch.h>
+#include <ATen/ops/add_cpu_dispatch.h>
+#include <ATen/ops/addbmm_cpu_dispatch.h>
+#include <ATen/ops/addcdiv_cpu_dispatch.h>
+#include <ATen/ops/addcmul_cpu_dispatch.h>
+#include <ATen/ops/addmm_cpu_dispatch.h>
+#include <ATen/ops/addmv_cpu_dispatch.h>
+#include <ATen/ops/addr_cpu_dispatch.h>
+#include <ATen/ops/all_cpu_dispatch.h>
+#include <ATen/ops/amax_cpu_dispatch.h>
+#include <ATen/ops/amin_cpu_dispatch.h>
+#include <ATen/ops/aminmax_cpu_dispatch.h>
+#include <ATen/ops/angle_cpu_dispatch.h>
+#include <ATen/ops/any_cpu_dispatch.h>
+#include <ATen/ops/arange_cpu_dispatch.h>
+#include <ATen/ops/argmax_cpu_dispatch.h>
+#include <ATen/ops/argmin_cpu_dispatch.h>
+#include <ATen/ops/argsort_cpu_dispatch.h>
+#include <ATen/ops/as_strided_cpu_dispatch.h>
+#include <ATen/ops/asin_cpu_dispatch.h>
+#include <ATen/ops/asinh_cpu_dispatch.h>
+#include <ATen/ops/atan_cpu_dispatch.h>
+#include <ATen/ops/atan2_cpu_dispatch.h>
+#include <ATen/ops/atanh_cpu_dispatch.h>
+#include <ATen/ops/avg_pool2d_cpu_dispatch.h>
+#include <ATen/ops/avg_pool2d_backward_cpu_dispatch.h>
+#include <ATen/ops/avg_pool3d_cpu_dispatch.h>
+#include <ATen/ops/avg_pool3d_backward_cpu_dispatch.h>
+#include <ATen/ops/baddbmm_cpu_dispatch.h>
+#include <ATen/ops/batch_norm_update_stats_cpu_dispatch.h>
+#include <ATen/ops/bernoulli_cpu_dispatch.h>
+#include <ATen/ops/binary_cross_entropy_cpu_dispatch.h>
+#include <ATen/ops/binary_cross_entropy_backward_cpu_dispatch.h>
+#include <ATen/ops/bincount_cpu_dispatch.h>
+#include <ATen/ops/binomial_cpu_dispatch.h>
+#include <ATen/ops/bitwise_and_cpu_dispatch.h>
+#include <ATen/ops/bitwise_left_shift_cpu_dispatch.h>
+#include <ATen/ops/bitwise_not_cpu_dispatch.h>
+#include <ATen/ops/bitwise_or_cpu_dispatch.h>
+#include <ATen/ops/bitwise_right_shift_cpu_dispatch.h>
+#include <ATen/ops/bitwise_xor_cpu_dispatch.h>
+#include <ATen/ops/bmm_cpu_dispatch.h>
+#include <ATen/ops/bucketize_cpu_dispatch.h>
+#include <ATen/ops/cat_cpu_dispatch.h>
+#include <ATen/ops/cauchy_cpu_dispatch.h>
+#include <ATen/ops/ceil_cpu_dispatch.h>
+#include <ATen/ops/channel_shuffle_cpu_dispatch.h>
+#include <ATen/ops/cholesky_cpu_dispatch.h>
+#include <ATen/ops/cholesky_inverse_cpu_dispatch.h>
+#include <ATen/ops/clamp_cpu_dispatch.h>
+#include <ATen/ops/clamp_max_cpu_dispatch.h>
+#include <ATen/ops/clamp_min_cpu_dispatch.h>
+#include <ATen/ops/col2im_cpu_dispatch.h>
+#include <ATen/ops/complex_cpu_dispatch.h>
+#include <ATen/ops/conj_physical_cpu_dispatch.h>
+#include <ATen/ops/copysign_cpu_dispatch.h>
+#include <ATen/ops/cos_cpu_dispatch.h>
+#include <ATen/ops/cosh_cpu_dispatch.h>
+#include <ATen/ops/count_nonzero_cpu_dispatch.h>
+#include <ATen/ops/cumprod_cpu_dispatch.h>
+#include <ATen/ops/cumsum_cpu_dispatch.h>
+#include <ATen/ops/dense_dim_cpu_dispatch.h>
+#include <ATen/ops/dequantize_cpu_dispatch.h>
+#include <ATen/ops/digamma_cpu_dispatch.h>
+#include <ATen/ops/div_cpu_dispatch.h>
+#include <ATen/ops/dot_cpu_dispatch.h>
+#include <ATen/ops/elu_cpu_dispatch.h>
+#include <ATen/ops/elu_backward_cpu_dispatch.h>
+#include <ATen/ops/embedding_dense_backward_cpu_dispatch.h>
+#include <ATen/ops/embedding_renorm_cpu_dispatch.h>
+#include <ATen/ops/empty_cpu_dispatch.h>
+#include <ATen/ops/empty_strided_cpu_dispatch.h>
+#include <ATen/ops/eq_cpu_dispatch.h>
+#include <ATen/ops/equal_cpu_dispatch.h>
+#include <ATen/ops/erf_cpu_dispatch.h>
+#include <ATen/ops/erfc_cpu_dispatch.h>
+#include <ATen/ops/erfinv_cpu_dispatch.h>
+#include <ATen/ops/exp_cpu_dispatch.h>
+#include <ATen/ops/exp2_cpu_dispatch.h>
+#include <ATen/ops/expm1_cpu_dispatch.h>
+#include <ATen/ops/exponential_cpu_dispatch.h>
+#include <ATen/ops/eye_cpu_dispatch.h>
+#include <ATen/ops/fake_quantize_per_channel_affine_cachemask_cpu_dispatch.h>
+#include <ATen/ops/fake_quantize_per_tensor_affine_cachemask_cpu_dispatch.h>
+#include <ATen/ops/fill_cpu_dispatch.h>
+#include <ATen/ops/flip_cpu_dispatch.h>
+#include <ATen/ops/floor_cpu_dispatch.h>
+#include <ATen/ops/floor_divide_cpu_dispatch.h>
+#include <ATen/ops/fmax_cpu_dispatch.h>
+#include <ATen/ops/fmin_cpu_dispatch.h>
+#include <ATen/ops/fmod_cpu_dispatch.h>
+#include <ATen/ops/frac_cpu_dispatch.h>
+#include <ATen/ops/fractional_max_pool2d_cpu_dispatch.h>
+#include <ATen/ops/fractional_max_pool2d_backward_cpu_dispatch.h>
+#include <ATen/ops/fractional_max_pool3d_cpu_dispatch.h>
+#include <ATen/ops/fractional_max_pool3d_backward_cpu_dispatch.h>
+#include <ATen/ops/frexp_cpu_dispatch.h>
+#include <ATen/ops/from_file_cpu_dispatch.h>
+#include <ATen/ops/gather_cpu_dispatch.h>
+#include <ATen/ops/gcd_cpu_dispatch.h>
+#include <ATen/ops/ge_cpu_dispatch.h>
+#include <ATen/ops/gelu_cpu_dispatch.h>
+#include <ATen/ops/gelu_backward_cpu_dispatch.h>
+#include <ATen/ops/geometric_cpu_dispatch.h>
+#include <ATen/ops/geqrf_cpu_dispatch.h>
+#include <ATen/ops/glu_cpu_dispatch.h>
+#include <ATen/ops/glu_backward_cpu_dispatch.h>
+#include <ATen/ops/glu_backward_jvp_cpu_dispatch.h>
+#include <ATen/ops/glu_jvp_cpu_dispatch.h>
+#include <ATen/ops/grid_sampler_2d_cpu_dispatch.h>
+#include <ATen/ops/grid_sampler_2d_backward_cpu_dispatch.h>
+#include <ATen/ops/grid_sampler_3d_cpu_dispatch.h>
+#include <ATen/ops/grid_sampler_3d_backward_cpu_dispatch.h>
+#include <ATen/ops/gt_cpu_dispatch.h>
+#include <ATen/ops/hardshrink_cpu_dispatch.h>
+#include <ATen/ops/hardshrink_backward_cpu_dispatch.h>
+#include <ATen/ops/hardsigmoid_cpu_dispatch.h>
+#include <ATen/ops/hardsigmoid_backward_cpu_dispatch.h>
+#include <ATen/ops/hardswish_cpu_dispatch.h>
+#include <ATen/ops/hardswish_backward_cpu_dispatch.h>
+#include <ATen/ops/hardtanh_cpu_dispatch.h>
+#include <ATen/ops/hardtanh_backward_cpu_dispatch.h>
+#include <ATen/ops/heaviside_cpu_dispatch.h>
+#include <ATen/ops/histc_cpu_dispatch.h>
+#include <ATen/ops/histogram_cpu_dispatch.h>
+#include <ATen/ops/huber_loss_cpu_dispatch.h>
+#include <ATen/ops/huber_loss_backward_cpu_dispatch.h>
+#include <ATen/ops/hypot_cpu_dispatch.h>
+#include <ATen/ops/i0_cpu_dispatch.h>
+#include <ATen/ops/igamma_cpu_dispatch.h>
+#include <ATen/ops/igammac_cpu_dispatch.h>
+#include <ATen/ops/im2col_cpu_dispatch.h>
+#include <ATen/ops/index_cpu_dispatch.h>
+#include <ATen/ops/index_add_cpu_dispatch.h>
+#include <ATen/ops/index_copy_cpu_dispatch.h>
+#include <ATen/ops/index_fill_cpu_dispatch.h>
+#include <ATen/ops/index_reduce_cpu_dispatch.h>
+#include <ATen/ops/index_select_cpu_dispatch.h>
+#include <ATen/ops/is_set_to_cpu_dispatch.h>
+#include <ATen/ops/isin_cpu_dispatch.h>
+#include <ATen/ops/isnan_cpu_dispatch.h>
+#include <ATen/ops/isneginf_cpu_dispatch.h>
+#include <ATen/ops/isposinf_cpu_dispatch.h>
+#include <ATen/ops/kthvalue_cpu_dispatch.h>
+#include <ATen/ops/lcm_cpu_dispatch.h>
+#include <ATen/ops/le_cpu_dispatch.h>
+#include <ATen/ops/leaky_relu_cpu_dispatch.h>
+#include <ATen/ops/leaky_relu_backward_cpu_dispatch.h>
+#include <ATen/ops/lerp_cpu_dispatch.h>
+#include <ATen/ops/lgamma_cpu_dispatch.h>
+#include <ATen/ops/linalg_cholesky_ex_cpu_dispatch.h>
+#include <ATen/ops/linalg_cross_cpu_dispatch.h>
+#include <ATen/ops/linalg_eig_cpu_dispatch.h>
+#include <ATen/ops/linalg_eigvals_cpu_dispatch.h>
+#include <ATen/ops/linalg_householder_product_cpu_dispatch.h>
+#include <ATen/ops/linalg_inv_ex_cpu_dispatch.h>
+#include <ATen/ops/linalg_ldl_factor_ex_cpu_dispatch.h>
+#include <ATen/ops/linalg_ldl_solve_cpu_dispatch.h>
+#include <ATen/ops/linalg_lstsq_cpu_dispatch.h>
+#include <ATen/ops/linalg_lu_cpu_dispatch.h>
+#include <ATen/ops/linalg_lu_factor_ex_cpu_dispatch.h>
+#include <ATen/ops/linalg_lu_solve_cpu_dispatch.h>
+#include <ATen/ops/linalg_matrix_exp_cpu_dispatch.h>
+#include <ATen/ops/linalg_qr_cpu_dispatch.h>
+#include <ATen/ops/linalg_solve_triangular_cpu_dispatch.h>
+#include <ATen/ops/linalg_vector_norm_cpu_dispatch.h>
+#include <ATen/ops/linspace_cpu_dispatch.h>
+#include <ATen/ops/log_cpu_dispatch.h>
+#include <ATen/ops/log10_cpu_dispatch.h>
+#include <ATen/ops/log1p_cpu_dispatch.h>
+#include <ATen/ops/log2_cpu_dispatch.h>
+#include <ATen/ops/log_normal_cpu_dispatch.h>
+#include <ATen/ops/log_sigmoid_backward_cpu_dispatch.h>
+#include <ATen/ops/log_sigmoid_forward_cpu_dispatch.h>
+#include <ATen/ops/logaddexp_cpu_dispatch.h>
+#include <ATen/ops/logaddexp2_cpu_dispatch.h>
+#include <ATen/ops/logical_and_cpu_dispatch.h>
+#include <ATen/ops/logical_not_cpu_dispatch.h>
+#include <ATen/ops/logical_or_cpu_dispatch.h>
+#include <ATen/ops/logical_xor_cpu_dispatch.h>
+#include <ATen/ops/logit_cpu_dispatch.h>
+#include <ATen/ops/logit_backward_cpu_dispatch.h>
+#include <ATen/ops/logspace_cpu_dispatch.h>
+#include <ATen/ops/lshift_cpu_dispatch.h>
+#include <ATen/ops/lt_cpu_dispatch.h>
+#include <ATen/ops/lu_unpack_cpu_dispatch.h>
+#include <ATen/ops/masked_fill_cpu_dispatch.h>
+#include <ATen/ops/masked_scatter_cpu_dispatch.h>
+#include <ATen/ops/masked_select_cpu_dispatch.h>
+#include <ATen/ops/max_cpu_dispatch.h>
+#include <ATen/ops/max_pool2d_with_indices_cpu_dispatch.h>
+#include <ATen/ops/max_pool2d_with_indices_backward_cpu_dispatch.h>
+#include <ATen/ops/max_pool3d_with_indices_cpu_dispatch.h>
+#include <ATen/ops/max_pool3d_with_indices_backward_cpu_dispatch.h>
+#include <ATen/ops/max_unpool2d_cpu_dispatch.h>
+#include <ATen/ops/max_unpool3d_cpu_dispatch.h>
+#include <ATen/ops/maximum_cpu_dispatch.h>
+#include <ATen/ops/mean_cpu_dispatch.h>
+#include <ATen/ops/median_cpu_dispatch.h>
+#include <ATen/ops/min_cpu_dispatch.h>
+#include <ATen/ops/minimum_cpu_dispatch.h>
+#include <ATen/ops/mish_cpu_dispatch.h>
+#include <ATen/ops/mish_backward_cpu_dispatch.h>
+#include <ATen/ops/mkldnn_rnn_layer_cpu_dispatch.h>
+#include <ATen/ops/mkldnn_rnn_layer_backward_cpu_dispatch.h>
+#include <ATen/ops/mm_cpu_dispatch.h>
+#include <ATen/ops/mode_cpu_dispatch.h>
+#include <ATen/ops/mse_loss_cpu_dispatch.h>
+#include <ATen/ops/mse_loss_backward_cpu_dispatch.h>
+#include <ATen/ops/mul_cpu_dispatch.h>
+#include <ATen/ops/multi_margin_loss_cpu_dispatch.h>
+#include <ATen/ops/multi_margin_loss_backward_cpu_dispatch.h>
+#include <ATen/ops/multilabel_margin_loss_backward_cpu_dispatch.h>
+#include <ATen/ops/multilabel_margin_loss_forward_cpu_dispatch.h>
+#include <ATen/ops/multinomial_cpu_dispatch.h>
+#include <ATen/ops/mvlgamma_cpu_dispatch.h>
+#include <ATen/ops/nan_to_num_cpu_dispatch.h>
+#include <ATen/ops/nanmedian_cpu_dispatch.h>
+#include <ATen/ops/nansum_cpu_dispatch.h>
+#include <ATen/ops/narrow_copy_cpu_dispatch.h>
+#include <ATen/ops/native_batch_norm_cpu_dispatch.h>
+#include <ATen/ops/native_batch_norm_backward_cpu_dispatch.h>
+#include <ATen/ops/native_channel_shuffle_cpu_dispatch.h>
+#include <ATen/ops/native_dropout_cpu_dispatch.h>
+#include <ATen/ops/native_dropout_backward_cpu_dispatch.h>
+#include <ATen/ops/native_group_norm_cpu_dispatch.h>
+#include <ATen/ops/native_group_norm_backward_cpu_dispatch.h>
+#include <ATen/ops/native_layer_norm_cpu_dispatch.h>
+#include <ATen/ops/native_layer_norm_backward_cpu_dispatch.h>
+#include <ATen/ops/ne_cpu_dispatch.h>
+#include <ATen/ops/neg_cpu_dispatch.h>
+#include <ATen/ops/nextafter_cpu_dispatch.h>
+#include <ATen/ops/nll_loss2d_backward_cpu_dispatch.h>
+#include <ATen/ops/nll_loss2d_forward_cpu_dispatch.h>
+#include <ATen/ops/nll_loss_backward_cpu_dispatch.h>
+#include <ATen/ops/nll_loss_forward_cpu_dispatch.h>
+#include <ATen/ops/nonzero_cpu_dispatch.h>
+#include <ATen/ops/nonzero_static_cpu_dispatch.h>
+#include <ATen/ops/norm_cpu_dispatch.h>
+#include <ATen/ops/normal_cpu_dispatch.h>
+#include <ATen/ops/ormqr_cpu_dispatch.h>
+#include <ATen/ops/pixel_shuffle_cpu_dispatch.h>
+#include <ATen/ops/pixel_unshuffle_cpu_dispatch.h>
+#include <ATen/ops/poisson_cpu_dispatch.h>
+#include <ATen/ops/polar_cpu_dispatch.h>
+#include <ATen/ops/polygamma_cpu_dispatch.h>
+#include <ATen/ops/pow_cpu_dispatch.h>
+#include <ATen/ops/prod_cpu_dispatch.h>
+#include <ATen/ops/put_cpu_dispatch.h>
+#include <ATen/ops/quantize_per_channel_cpu_dispatch.h>
+#include <ATen/ops/quantize_per_tensor_cpu_dispatch.h>
+#include <ATen/ops/quantize_per_tensor_dynamic_cpu_dispatch.h>
+#include <ATen/ops/random_cpu_dispatch.h>
+#include <ATen/ops/randperm_cpu_dispatch.h>
+#include <ATen/ops/range_cpu_dispatch.h>
+#include <ATen/ops/reciprocal_cpu_dispatch.h>
+#include <ATen/ops/reflection_pad1d_cpu_dispatch.h>
+#include <ATen/ops/reflection_pad1d_backward_cpu_dispatch.h>
+#include <ATen/ops/reflection_pad2d_cpu_dispatch.h>
+#include <ATen/ops/reflection_pad2d_backward_cpu_dispatch.h>
+#include <ATen/ops/reflection_pad3d_cpu_dispatch.h>
+#include <ATen/ops/reflection_pad3d_backward_cpu_dispatch.h>
+#include <ATen/ops/relu_cpu_dispatch.h>
+#include <ATen/ops/remainder_cpu_dispatch.h>
+#include <ATen/ops/renorm_cpu_dispatch.h>
+#include <ATen/ops/repeat_interleave_cpu_dispatch.h>
+#include <ATen/ops/replication_pad1d_cpu_dispatch.h>
+#include <ATen/ops/replication_pad1d_backward_cpu_dispatch.h>
+#include <ATen/ops/replication_pad2d_cpu_dispatch.h>
+#include <ATen/ops/replication_pad2d_backward_cpu_dispatch.h>
+#include <ATen/ops/replication_pad3d_cpu_dispatch.h>
+#include <ATen/ops/replication_pad3d_backward_cpu_dispatch.h>
+#include <ATen/ops/resize_cpu_dispatch.h>
+#include <ATen/ops/roll_cpu_dispatch.h>
+#include <ATen/ops/round_cpu_dispatch.h>
+#include <ATen/ops/rrelu_with_noise_cpu_dispatch.h>
+#include <ATen/ops/rshift_cpu_dispatch.h>
+#include <ATen/ops/rsqrt_cpu_dispatch.h>
+#include <ATen/ops/rsub_cpu_dispatch.h>
+#include <ATen/ops/scatter_cpu_dispatch.h>
+#include <ATen/ops/scatter_add_cpu_dispatch.h>
+#include <ATen/ops/scatter_reduce_cpu_dispatch.h>
+#include <ATen/ops/searchsorted_cpu_dispatch.h>
+#include <ATen/ops/segment_reduce_cpu_dispatch.h>
+#include <ATen/ops/set_cpu_dispatch.h>
+#include <ATen/ops/sgn_cpu_dispatch.h>
+#include <ATen/ops/sigmoid_cpu_dispatch.h>
+#include <ATen/ops/sigmoid_backward_cpu_dispatch.h>
+#include <ATen/ops/sign_cpu_dispatch.h>
+#include <ATen/ops/signbit_cpu_dispatch.h>
+#include <ATen/ops/silu_cpu_dispatch.h>
+#include <ATen/ops/silu_backward_cpu_dispatch.h>
+#include <ATen/ops/sin_cpu_dispatch.h>
+#include <ATen/ops/sinc_cpu_dispatch.h>
+#include <ATen/ops/sinh_cpu_dispatch.h>
+#include <ATen/ops/slow_conv3d_forward_cpu_dispatch.h>
+#include <ATen/ops/slow_conv_dilated2d_cpu_dispatch.h>
+#include <ATen/ops/slow_conv_dilated3d_cpu_dispatch.h>
+#include <ATen/ops/slow_conv_transpose2d_cpu_dispatch.h>
+#include <ATen/ops/slow_conv_transpose3d_cpu_dispatch.h>
+#include <ATen/ops/smooth_l1_loss_cpu_dispatch.h>
+#include <ATen/ops/smooth_l1_loss_backward_cpu_dispatch.h>
+#include <ATen/ops/softplus_cpu_dispatch.h>
+#include <ATen/ops/softplus_backward_cpu_dispatch.h>
+#include <ATen/ops/softshrink_cpu_dispatch.h>
+#include <ATen/ops/softshrink_backward_cpu_dispatch.h>
+#include <ATen/ops/sort_cpu_dispatch.h>
+#include <ATen/ops/sparse_dim_cpu_dispatch.h>
+#include <ATen/ops/special_airy_ai_cpu_dispatch.h>
+#include <ATen/ops/special_bessel_j0_cpu_dispatch.h>
+#include <ATen/ops/special_bessel_j1_cpu_dispatch.h>
+#include <ATen/ops/special_bessel_y0_cpu_dispatch.h>
+#include <ATen/ops/special_bessel_y1_cpu_dispatch.h>
+#include <ATen/ops/special_chebyshev_polynomial_t_cpu_dispatch.h>
+#include <ATen/ops/special_chebyshev_polynomial_u_cpu_dispatch.h>
+#include <ATen/ops/special_chebyshev_polynomial_v_cpu_dispatch.h>
+#include <ATen/ops/special_chebyshev_polynomial_w_cpu_dispatch.h>
+#include <ATen/ops/special_entr_cpu_dispatch.h>
+#include <ATen/ops/special_erfcx_cpu_dispatch.h>
+#include <ATen/ops/special_hermite_polynomial_h_cpu_dispatch.h>
+#include <ATen/ops/special_hermite_polynomial_he_cpu_dispatch.h>
+#include <ATen/ops/special_i0e_cpu_dispatch.h>
+#include <ATen/ops/special_i1_cpu_dispatch.h>
+#include <ATen/ops/special_i1e_cpu_dispatch.h>
+#include <ATen/ops/special_laguerre_polynomial_l_cpu_dispatch.h>
+#include <ATen/ops/special_legendre_polynomial_p_cpu_dispatch.h>
+#include <ATen/ops/special_log_ndtr_cpu_dispatch.h>
+#include <ATen/ops/special_modified_bessel_i0_cpu_dispatch.h>
+#include <ATen/ops/special_modified_bessel_i1_cpu_dispatch.h>
+#include <ATen/ops/special_modified_bessel_k0_cpu_dispatch.h>
+#include <ATen/ops/special_modified_bessel_k1_cpu_dispatch.h>
+#include <ATen/ops/special_ndtri_cpu_dispatch.h>
+#include <ATen/ops/special_scaled_modified_bessel_k0_cpu_dispatch.h>
+#include <ATen/ops/special_scaled_modified_bessel_k1_cpu_dispatch.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_t_cpu_dispatch.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_u_cpu_dispatch.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_v_cpu_dispatch.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_w_cpu_dispatch.h>
+#include <ATen/ops/special_spherical_bessel_j0_cpu_dispatch.h>
+#include <ATen/ops/special_xlog1py_cpu_dispatch.h>
+#include <ATen/ops/special_zeta_cpu_dispatch.h>
+#include <ATen/ops/sqrt_cpu_dispatch.h>
+#include <ATen/ops/sspaddmm_cpu_dispatch.h>
+#include <ATen/ops/std_cpu_dispatch.h>
+#include <ATen/ops/std_mean_cpu_dispatch.h>
+#include <ATen/ops/sub_cpu_dispatch.h>
+#include <ATen/ops/sum_cpu_dispatch.h>
+#include <ATen/ops/take_cpu_dispatch.h>
+#include <ATen/ops/tan_cpu_dispatch.h>
+#include <ATen/ops/tanh_cpu_dispatch.h>
+#include <ATen/ops/tanh_backward_cpu_dispatch.h>
+#include <ATen/ops/threshold_cpu_dispatch.h>
+#include <ATen/ops/threshold_backward_cpu_dispatch.h>
+#include <ATen/ops/to_mkldnn_cpu_dispatch.h>
+#include <ATen/ops/topk_cpu_dispatch.h>
+#include <ATen/ops/trace_cpu_dispatch.h>
+#include <ATen/ops/triangular_solve_cpu_dispatch.h>
+#include <ATen/ops/tril_cpu_dispatch.h>
+#include <ATen/ops/tril_indices_cpu_dispatch.h>
+#include <ATen/ops/triu_cpu_dispatch.h>
+#include <ATen/ops/triu_indices_cpu_dispatch.h>
+#include <ATen/ops/trunc_cpu_dispatch.h>
+#include <ATen/ops/unfold_cpu_dispatch.h>
+#include <ATen/ops/unfold_backward_cpu_dispatch.h>
+#include <ATen/ops/uniform_cpu_dispatch.h>
+#include <ATen/ops/unique_consecutive_cpu_dispatch.h>
+#include <ATen/ops/unique_dim_cpu_dispatch.h>
+#include <ATen/ops/unique_dim_consecutive_cpu_dispatch.h>
+#include <ATen/ops/upsample_bicubic2d_cpu_dispatch.h>
+#include <ATen/ops/upsample_bicubic2d_backward_cpu_dispatch.h>
+#include <ATen/ops/upsample_bilinear2d_cpu_dispatch.h>
+#include <ATen/ops/upsample_bilinear2d_backward_cpu_dispatch.h>
+#include <ATen/ops/upsample_linear1d_cpu_dispatch.h>
+#include <ATen/ops/upsample_linear1d_backward_cpu_dispatch.h>
+#include <ATen/ops/upsample_nearest1d_cpu_dispatch.h>
+#include <ATen/ops/upsample_nearest1d_backward_cpu_dispatch.h>
+#include <ATen/ops/upsample_nearest2d_cpu_dispatch.h>
+#include <ATen/ops/upsample_nearest2d_backward_cpu_dispatch.h>
+#include <ATen/ops/upsample_nearest3d_cpu_dispatch.h>
+#include <ATen/ops/upsample_nearest3d_backward_cpu_dispatch.h>
+#include <ATen/ops/upsample_trilinear3d_cpu_dispatch.h>
+#include <ATen/ops/upsample_trilinear3d_backward_cpu_dispatch.h>
+#include <ATen/ops/var_cpu_dispatch.h>
+#include <ATen/ops/var_mean_cpu_dispatch.h>
+#include <ATen/ops/vdot_cpu_dispatch.h>
+#include <ATen/ops/view_cpu_dispatch.h>
+#include <ATen/ops/view_as_complex_cpu_dispatch.h>
+#include <ATen/ops/view_as_real_cpu_dispatch.h>
+#include <ATen/ops/where_cpu_dispatch.h>
+#include <ATen/ops/xlogy_cpu_dispatch.h>
+#include <ATen/ops/zero_cpu_dispatch.h>
+
+
+

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

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

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/CUDAFunctions.h

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

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/CUDAFunctions_inl.h

@@ -0,0 +1,614 @@
+#pragma once
+// @generated by torchgen/gen.py from DispatchKeyFunctions_inl.h
+
+// NB: The implementing C++ file is RegisterDispatchKey.cpp
+
+// The only #includes we need are for custom classes that have defaults in the C++ API
+#include <c10/core/MemoryFormat.h>
+#include <c10/core/Scalar.h>
+#include <ATen/core/Reduction.h>
+
+#if defined(AT_PER_OPERATOR_HEADERS) && defined(TORCH_ASSERT_ONLY_METHOD_OPERATORS)
+#error This change adds a dependency on all pytorch operators, meaning the     \
+  file will need to be re-compiled every time an operator is changed or added. \
+  Consider including a specific operator from                                  \
+  <ATen/ops/{my_operator}_cuda_dispatch.h>.                   \
+  See NOTE [TORCH_ASSERT_ONLY_METHOD_OPERATORS].
+#endif
+
+#include <ATen/ops/_adaptive_avg_pool2d_cuda_dispatch.h>
+#include <ATen/ops/_adaptive_avg_pool2d_backward_cuda_dispatch.h>
+#include <ATen/ops/_adaptive_avg_pool3d_cuda_dispatch.h>
+#include <ATen/ops/_adaptive_avg_pool3d_backward_cuda_dispatch.h>
+#include <ATen/ops/_addmm_activation_cuda_dispatch.h>
+#include <ATen/ops/_aminmax_cuda_dispatch.h>
+#include <ATen/ops/_amp_foreach_non_finite_check_and_unscale_cuda_dispatch.h>
+#include <ATen/ops/_amp_update_scale_cuda_dispatch.h>
+#include <ATen/ops/_assert_async_cuda_dispatch.h>
+#include <ATen/ops/_cdist_backward_cuda_dispatch.h>
+#include <ATen/ops/_cdist_forward_cuda_dispatch.h>
+#include <ATen/ops/_cholesky_solve_helper_cuda_dispatch.h>
+#include <ATen/ops/_compute_linear_combination_cuda_dispatch.h>
+#include <ATen/ops/_conv_depthwise2d_cuda_dispatch.h>
+#include <ATen/ops/_convert_indices_from_coo_to_csr_cuda_dispatch.h>
+#include <ATen/ops/_convert_indices_from_csr_to_coo_cuda_dispatch.h>
+#include <ATen/ops/_convert_weight_to_int4pack_cuda_dispatch.h>
+#include <ATen/ops/_cslt_compress_cuda_dispatch.h>
+#include <ATen/ops/_cslt_sparse_mm_cuda_dispatch.h>
+#include <ATen/ops/_cslt_sparse_mm_search_cuda_dispatch.h>
+#include <ATen/ops/_ctc_loss_cuda_dispatch.h>
+#include <ATen/ops/_ctc_loss_backward_cuda_dispatch.h>
+#include <ATen/ops/_cudnn_ctc_loss_cuda_dispatch.h>
+#include <ATen/ops/_cudnn_init_dropout_state_cuda_dispatch.h>
+#include <ATen/ops/_cudnn_rnn_cuda_dispatch.h>
+#include <ATen/ops/_cudnn_rnn_backward_cuda_dispatch.h>
+#include <ATen/ops/_cudnn_rnn_flatten_weight_cuda_dispatch.h>
+#include <ATen/ops/_cummax_helper_cuda_dispatch.h>
+#include <ATen/ops/_cummin_helper_cuda_dispatch.h>
+#include <ATen/ops/_dirichlet_grad_cuda_dispatch.h>
+#include <ATen/ops/_efficient_attention_backward_cuda_dispatch.h>
+#include <ATen/ops/_efficient_attention_forward_cuda_dispatch.h>
+#include <ATen/ops/_efficientzerotensor_cuda_dispatch.h>
+#include <ATen/ops/_embedding_bag_cuda_dispatch.h>
+#include <ATen/ops/_embedding_bag_dense_backward_cuda_dispatch.h>
+#include <ATen/ops/_embedding_bag_forward_only_cuda_dispatch.h>
+#include <ATen/ops/_embedding_bag_per_sample_weights_backward_cuda_dispatch.h>
+#include <ATen/ops/_fake_quantize_learnable_per_channel_affine_cuda_dispatch.h>
+#include <ATen/ops/_fake_quantize_learnable_per_channel_affine_backward_cuda_dispatch.h>
+#include <ATen/ops/_fake_quantize_learnable_per_tensor_affine_cuda_dispatch.h>
+#include <ATen/ops/_fake_quantize_learnable_per_tensor_affine_backward_cuda_dispatch.h>
+#include <ATen/ops/_fake_quantize_per_tensor_affine_cachemask_tensor_qparams_cuda_dispatch.h>
+#include <ATen/ops/_fft_c2c_cuda_dispatch.h>
+#include <ATen/ops/_fft_c2r_cuda_dispatch.h>
+#include <ATen/ops/_fft_r2c_cuda_dispatch.h>
+#include <ATen/ops/_fill_mem_eff_dropout_mask_cuda_dispatch.h>
+#include <ATen/ops/_flash_attention_backward_cuda_dispatch.h>
+#include <ATen/ops/_flash_attention_forward_cuda_dispatch.h>
+#include <ATen/ops/_foreach_abs_cuda_dispatch.h>
+#include <ATen/ops/_foreach_acos_cuda_dispatch.h>
+#include <ATen/ops/_foreach_add_cuda_dispatch.h>
+#include <ATen/ops/_foreach_addcdiv_cuda_dispatch.h>
+#include <ATen/ops/_foreach_addcmul_cuda_dispatch.h>
+#include <ATen/ops/_foreach_asin_cuda_dispatch.h>
+#include <ATen/ops/_foreach_atan_cuda_dispatch.h>
+#include <ATen/ops/_foreach_ceil_cuda_dispatch.h>
+#include <ATen/ops/_foreach_clamp_max_cuda_dispatch.h>
+#include <ATen/ops/_foreach_clamp_min_cuda_dispatch.h>
+#include <ATen/ops/_foreach_copy_cuda_dispatch.h>
+#include <ATen/ops/_foreach_cos_cuda_dispatch.h>
+#include <ATen/ops/_foreach_cosh_cuda_dispatch.h>
+#include <ATen/ops/_foreach_div_cuda_dispatch.h>
+#include <ATen/ops/_foreach_erf_cuda_dispatch.h>
+#include <ATen/ops/_foreach_erfc_cuda_dispatch.h>
+#include <ATen/ops/_foreach_exp_cuda_dispatch.h>
+#include <ATen/ops/_foreach_expm1_cuda_dispatch.h>
+#include <ATen/ops/_foreach_floor_cuda_dispatch.h>
+#include <ATen/ops/_foreach_frac_cuda_dispatch.h>
+#include <ATen/ops/_foreach_lerp_cuda_dispatch.h>
+#include <ATen/ops/_foreach_lgamma_cuda_dispatch.h>
+#include <ATen/ops/_foreach_log_cuda_dispatch.h>
+#include <ATen/ops/_foreach_log10_cuda_dispatch.h>
+#include <ATen/ops/_foreach_log1p_cuda_dispatch.h>
+#include <ATen/ops/_foreach_log2_cuda_dispatch.h>
+#include <ATen/ops/_foreach_maximum_cuda_dispatch.h>
+#include <ATen/ops/_foreach_minimum_cuda_dispatch.h>
+#include <ATen/ops/_foreach_mul_cuda_dispatch.h>
+#include <ATen/ops/_foreach_neg_cuda_dispatch.h>
+#include <ATen/ops/_foreach_norm_cuda_dispatch.h>
+#include <ATen/ops/_foreach_pow_cuda_dispatch.h>
+#include <ATen/ops/_foreach_reciprocal_cuda_dispatch.h>
+#include <ATen/ops/_foreach_round_cuda_dispatch.h>
+#include <ATen/ops/_foreach_sigmoid_cuda_dispatch.h>
+#include <ATen/ops/_foreach_sign_cuda_dispatch.h>
+#include <ATen/ops/_foreach_sin_cuda_dispatch.h>
+#include <ATen/ops/_foreach_sinh_cuda_dispatch.h>
+#include <ATen/ops/_foreach_sqrt_cuda_dispatch.h>
+#include <ATen/ops/_foreach_sub_cuda_dispatch.h>
+#include <ATen/ops/_foreach_tan_cuda_dispatch.h>
+#include <ATen/ops/_foreach_tanh_cuda_dispatch.h>
+#include <ATen/ops/_foreach_trunc_cuda_dispatch.h>
+#include <ATen/ops/_foreach_zero_cuda_dispatch.h>
+#include <ATen/ops/_fused_adam_cuda_dispatch.h>
+#include <ATen/ops/_fused_adamw_cuda_dispatch.h>
+#include <ATen/ops/_fused_dropout_cuda_dispatch.h>
+#include <ATen/ops/_fused_moving_avg_obs_fq_helper_cuda_dispatch.h>
+#include <ATen/ops/_fused_sdp_choice_cuda_dispatch.h>
+#include <ATen/ops/_fused_sgd_cuda_dispatch.h>
+#include <ATen/ops/_index_put_impl_cuda_dispatch.h>
+#include <ATen/ops/_int_mm_cuda_dispatch.h>
+#include <ATen/ops/_linalg_det_cuda_dispatch.h>
+#include <ATen/ops/_linalg_eigh_cuda_dispatch.h>
+#include <ATen/ops/_linalg_eigvals_cuda_dispatch.h>
+#include <ATen/ops/_linalg_slogdet_cuda_dispatch.h>
+#include <ATen/ops/_linalg_solve_ex_cuda_dispatch.h>
+#include <ATen/ops/_linalg_svd_cuda_dispatch.h>
+#include <ATen/ops/_local_scalar_dense_cuda_dispatch.h>
+#include <ATen/ops/_log_softmax_cuda_dispatch.h>
+#include <ATen/ops/_log_softmax_backward_data_cuda_dispatch.h>
+#include <ATen/ops/_logcumsumexp_cuda_dispatch.h>
+#include <ATen/ops/_make_per_channel_quantized_tensor_cuda_dispatch.h>
+#include <ATen/ops/_make_per_tensor_quantized_tensor_cuda_dispatch.h>
+#include <ATen/ops/_masked_scale_cuda_dispatch.h>
+#include <ATen/ops/_masked_softmax_cuda_dispatch.h>
+#include <ATen/ops/_masked_softmax_backward_cuda_dispatch.h>
+#include <ATen/ops/_mixed_dtypes_linear_cuda_dispatch.h>
+#include <ATen/ops/_native_batch_norm_legit_cuda_dispatch.h>
+#include <ATen/ops/_native_multi_head_attention_cuda_dispatch.h>
+#include <ATen/ops/_nested_from_padded_cuda_dispatch.h>
+#include <ATen/ops/_nested_tensor_from_mask_cuda_dispatch.h>
+#include <ATen/ops/_nested_tensor_from_mask_left_aligned_cuda_dispatch.h>
+#include <ATen/ops/_nested_view_from_buffer_cuda_dispatch.h>
+#include <ATen/ops/_pdist_backward_cuda_dispatch.h>
+#include <ATen/ops/_pdist_forward_cuda_dispatch.h>
+#include <ATen/ops/_pin_memory_cuda_dispatch.h>
+#include <ATen/ops/_prelu_kernel_cuda_dispatch.h>
+#include <ATen/ops/_prelu_kernel_backward_cuda_dispatch.h>
+#include <ATen/ops/_reshape_alias_cuda_dispatch.h>
+#include <ATen/ops/_sample_dirichlet_cuda_dispatch.h>
+#include <ATen/ops/_scaled_dot_product_cudnn_attention_cuda_dispatch.h>
+#include <ATen/ops/_scaled_dot_product_efficient_attention_cuda_dispatch.h>
+#include <ATen/ops/_scaled_dot_product_efficient_attention_backward_cuda_dispatch.h>
+#include <ATen/ops/_scaled_dot_product_flash_attention_cuda_dispatch.h>
+#include <ATen/ops/_scaled_dot_product_flash_attention_backward_cuda_dispatch.h>
+#include <ATen/ops/_scaled_mm_cuda_dispatch.h>
+#include <ATen/ops/_segment_reduce_backward_cuda_dispatch.h>
+#include <ATen/ops/_slow_conv2d_backward_cuda_dispatch.h>
+#include <ATen/ops/_slow_conv2d_forward_cuda_dispatch.h>
+#include <ATen/ops/_softmax_cuda_dispatch.h>
+#include <ATen/ops/_softmax_backward_data_cuda_dispatch.h>
+#include <ATen/ops/_sparse_semi_structured_linear_cuda_dispatch.h>
+#include <ATen/ops/_standard_gamma_cuda_dispatch.h>
+#include <ATen/ops/_standard_gamma_grad_cuda_dispatch.h>
+#include <ATen/ops/_thnn_fused_gru_cell_cuda_dispatch.h>
+#include <ATen/ops/_thnn_fused_gru_cell_backward_cuda_dispatch.h>
+#include <ATen/ops/_thnn_fused_lstm_cell_cuda_dispatch.h>
+#include <ATen/ops/_thnn_fused_lstm_cell_backward_impl_cuda_dispatch.h>
+#include <ATen/ops/_to_sparse_cuda_dispatch.h>
+#include <ATen/ops/_to_sparse_bsc_cuda_dispatch.h>
+#include <ATen/ops/_to_sparse_bsr_cuda_dispatch.h>
+#include <ATen/ops/_to_sparse_csc_cuda_dispatch.h>
+#include <ATen/ops/_to_sparse_csr_cuda_dispatch.h>
+#include <ATen/ops/_to_sparse_semi_structured_cuda_dispatch.h>
+#include <ATen/ops/_transform_bias_rescale_qkv_cuda_dispatch.h>
+#include <ATen/ops/_transformer_encoder_layer_fwd_cuda_dispatch.h>
+#include <ATen/ops/_triton_multi_head_attention_cuda_dispatch.h>
+#include <ATen/ops/_triton_scaled_dot_attention_cuda_dispatch.h>
+#include <ATen/ops/_unique_cuda_dispatch.h>
+#include <ATen/ops/_unique2_cuda_dispatch.h>
+#include <ATen/ops/_upsample_bicubic2d_aa_cuda_dispatch.h>
+#include <ATen/ops/_upsample_bicubic2d_aa_backward_cuda_dispatch.h>
+#include <ATen/ops/_upsample_bilinear2d_aa_cuda_dispatch.h>
+#include <ATen/ops/_upsample_bilinear2d_aa_backward_cuda_dispatch.h>
+#include <ATen/ops/_upsample_nearest_exact1d_cuda_dispatch.h>
+#include <ATen/ops/_upsample_nearest_exact1d_backward_cuda_dispatch.h>
+#include <ATen/ops/_upsample_nearest_exact2d_cuda_dispatch.h>
+#include <ATen/ops/_upsample_nearest_exact2d_backward_cuda_dispatch.h>
+#include <ATen/ops/_upsample_nearest_exact3d_cuda_dispatch.h>
+#include <ATen/ops/_upsample_nearest_exact3d_backward_cuda_dispatch.h>
+#include <ATen/ops/_use_cudnn_ctc_loss_cuda_dispatch.h>
+#include <ATen/ops/_validate_compressed_sparse_indices_cuda_dispatch.h>
+#include <ATen/ops/_weight_int4pack_mm_cuda_dispatch.h>
+#include <ATen/ops/_weight_norm_interface_cuda_dispatch.h>
+#include <ATen/ops/_weight_norm_interface_backward_cuda_dispatch.h>
+#include <ATen/ops/abs_cuda_dispatch.h>
+#include <ATen/ops/acos_cuda_dispatch.h>
+#include <ATen/ops/acosh_cuda_dispatch.h>
+#include <ATen/ops/adaptive_avg_pool2d_cuda_dispatch.h>
+#include <ATen/ops/adaptive_avg_pool3d_cuda_dispatch.h>
+#include <ATen/ops/adaptive_avg_pool3d_backward_cuda_dispatch.h>
+#include <ATen/ops/adaptive_max_pool2d_cuda_dispatch.h>
+#include <ATen/ops/adaptive_max_pool2d_backward_cuda_dispatch.h>
+#include <ATen/ops/adaptive_max_pool3d_cuda_dispatch.h>
+#include <ATen/ops/adaptive_max_pool3d_backward_cuda_dispatch.h>
+#include <ATen/ops/add_cuda_dispatch.h>
+#include <ATen/ops/addbmm_cuda_dispatch.h>
+#include <ATen/ops/addcdiv_cuda_dispatch.h>
+#include <ATen/ops/addcmul_cuda_dispatch.h>
+#include <ATen/ops/addmm_cuda_dispatch.h>
+#include <ATen/ops/addmv_cuda_dispatch.h>
+#include <ATen/ops/addr_cuda_dispatch.h>
+#include <ATen/ops/all_cuda_dispatch.h>
+#include <ATen/ops/amax_cuda_dispatch.h>
+#include <ATen/ops/amin_cuda_dispatch.h>
+#include <ATen/ops/aminmax_cuda_dispatch.h>
+#include <ATen/ops/angle_cuda_dispatch.h>
+#include <ATen/ops/any_cuda_dispatch.h>
+#include <ATen/ops/arange_cuda_dispatch.h>
+#include <ATen/ops/argmax_cuda_dispatch.h>
+#include <ATen/ops/argmin_cuda_dispatch.h>
+#include <ATen/ops/argsort_cuda_dispatch.h>
+#include <ATen/ops/as_strided_cuda_dispatch.h>
+#include <ATen/ops/asin_cuda_dispatch.h>
+#include <ATen/ops/asinh_cuda_dispatch.h>
+#include <ATen/ops/atan_cuda_dispatch.h>
+#include <ATen/ops/atan2_cuda_dispatch.h>
+#include <ATen/ops/atanh_cuda_dispatch.h>
+#include <ATen/ops/avg_pool2d_cuda_dispatch.h>
+#include <ATen/ops/avg_pool2d_backward_cuda_dispatch.h>
+#include <ATen/ops/avg_pool3d_cuda_dispatch.h>
+#include <ATen/ops/avg_pool3d_backward_cuda_dispatch.h>
+#include <ATen/ops/baddbmm_cuda_dispatch.h>
+#include <ATen/ops/batch_norm_backward_elemt_cuda_dispatch.h>
+#include <ATen/ops/batch_norm_backward_reduce_cuda_dispatch.h>
+#include <ATen/ops/batch_norm_elemt_cuda_dispatch.h>
+#include <ATen/ops/batch_norm_gather_stats_cuda_dispatch.h>
+#include <ATen/ops/batch_norm_gather_stats_with_counts_cuda_dispatch.h>
+#include <ATen/ops/batch_norm_stats_cuda_dispatch.h>
+#include <ATen/ops/batch_norm_update_stats_cuda_dispatch.h>
+#include <ATen/ops/bernoulli_cuda_dispatch.h>
+#include <ATen/ops/binary_cross_entropy_cuda_dispatch.h>
+#include <ATen/ops/binary_cross_entropy_backward_cuda_dispatch.h>
+#include <ATen/ops/bincount_cuda_dispatch.h>
+#include <ATen/ops/binomial_cuda_dispatch.h>
+#include <ATen/ops/bitwise_and_cuda_dispatch.h>
+#include <ATen/ops/bitwise_left_shift_cuda_dispatch.h>
+#include <ATen/ops/bitwise_not_cuda_dispatch.h>
+#include <ATen/ops/bitwise_or_cuda_dispatch.h>
+#include <ATen/ops/bitwise_right_shift_cuda_dispatch.h>
+#include <ATen/ops/bitwise_xor_cuda_dispatch.h>
+#include <ATen/ops/bmm_cuda_dispatch.h>
+#include <ATen/ops/bucketize_cuda_dispatch.h>
+#include <ATen/ops/cat_cuda_dispatch.h>
+#include <ATen/ops/cauchy_cuda_dispatch.h>
+#include <ATen/ops/ceil_cuda_dispatch.h>
+#include <ATen/ops/channel_shuffle_cuda_dispatch.h>
+#include <ATen/ops/cholesky_cuda_dispatch.h>
+#include <ATen/ops/cholesky_inverse_cuda_dispatch.h>
+#include <ATen/ops/clamp_cuda_dispatch.h>
+#include <ATen/ops/clamp_max_cuda_dispatch.h>
+#include <ATen/ops/clamp_min_cuda_dispatch.h>
+#include <ATen/ops/col2im_cuda_dispatch.h>
+#include <ATen/ops/complex_cuda_dispatch.h>
+#include <ATen/ops/conj_physical_cuda_dispatch.h>
+#include <ATen/ops/conv_depthwise3d_cuda_dispatch.h>
+#include <ATen/ops/convolution_backward_cuda_dispatch.h>
+#include <ATen/ops/copysign_cuda_dispatch.h>
+#include <ATen/ops/cos_cuda_dispatch.h>
+#include <ATen/ops/cosh_cuda_dispatch.h>
+#include <ATen/ops/count_nonzero_cuda_dispatch.h>
+#include <ATen/ops/cudnn_affine_grid_generator_cuda_dispatch.h>
+#include <ATen/ops/cudnn_affine_grid_generator_backward_cuda_dispatch.h>
+#include <ATen/ops/cudnn_batch_norm_cuda_dispatch.h>
+#include <ATen/ops/cudnn_batch_norm_backward_cuda_dispatch.h>
+#include <ATen/ops/cudnn_convolution_cuda_dispatch.h>
+#include <ATen/ops/cudnn_convolution_add_relu_cuda_dispatch.h>
+#include <ATen/ops/cudnn_convolution_relu_cuda_dispatch.h>
+#include <ATen/ops/cudnn_convolution_transpose_cuda_dispatch.h>
+#include <ATen/ops/cudnn_grid_sampler_cuda_dispatch.h>
+#include <ATen/ops/cudnn_grid_sampler_backward_cuda_dispatch.h>
+#include <ATen/ops/cumprod_cuda_dispatch.h>
+#include <ATen/ops/cumsum_cuda_dispatch.h>
+#include <ATen/ops/dense_dim_cuda_dispatch.h>
+#include <ATen/ops/dequantize_cuda_dispatch.h>
+#include <ATen/ops/digamma_cuda_dispatch.h>
+#include <ATen/ops/div_cuda_dispatch.h>
+#include <ATen/ops/dot_cuda_dispatch.h>
+#include <ATen/ops/elu_cuda_dispatch.h>
+#include <ATen/ops/elu_backward_cuda_dispatch.h>
+#include <ATen/ops/embedding_dense_backward_cuda_dispatch.h>
+#include <ATen/ops/embedding_renorm_cuda_dispatch.h>
+#include <ATen/ops/empty_cuda_dispatch.h>
+#include <ATen/ops/empty_strided_cuda_dispatch.h>
+#include <ATen/ops/eq_cuda_dispatch.h>
+#include <ATen/ops/equal_cuda_dispatch.h>
+#include <ATen/ops/erf_cuda_dispatch.h>
+#include <ATen/ops/erfc_cuda_dispatch.h>
+#include <ATen/ops/erfinv_cuda_dispatch.h>
+#include <ATen/ops/exp_cuda_dispatch.h>
+#include <ATen/ops/exp2_cuda_dispatch.h>
+#include <ATen/ops/expm1_cuda_dispatch.h>
+#include <ATen/ops/exponential_cuda_dispatch.h>
+#include <ATen/ops/eye_cuda_dispatch.h>
+#include <ATen/ops/fake_quantize_per_channel_affine_cachemask_cuda_dispatch.h>
+#include <ATen/ops/fake_quantize_per_tensor_affine_cachemask_cuda_dispatch.h>
+#include <ATen/ops/fill_cuda_dispatch.h>
+#include <ATen/ops/flip_cuda_dispatch.h>
+#include <ATen/ops/floor_cuda_dispatch.h>
+#include <ATen/ops/floor_divide_cuda_dispatch.h>
+#include <ATen/ops/fmax_cuda_dispatch.h>
+#include <ATen/ops/fmin_cuda_dispatch.h>
+#include <ATen/ops/fmod_cuda_dispatch.h>
+#include <ATen/ops/frac_cuda_dispatch.h>
+#include <ATen/ops/fractional_max_pool2d_cuda_dispatch.h>
+#include <ATen/ops/fractional_max_pool2d_backward_cuda_dispatch.h>
+#include <ATen/ops/fractional_max_pool3d_cuda_dispatch.h>
+#include <ATen/ops/fractional_max_pool3d_backward_cuda_dispatch.h>
+#include <ATen/ops/frexp_cuda_dispatch.h>
+#include <ATen/ops/gather_cuda_dispatch.h>
+#include <ATen/ops/gcd_cuda_dispatch.h>
+#include <ATen/ops/ge_cuda_dispatch.h>
+#include <ATen/ops/gelu_cuda_dispatch.h>
+#include <ATen/ops/gelu_backward_cuda_dispatch.h>
+#include <ATen/ops/geometric_cuda_dispatch.h>
+#include <ATen/ops/geqrf_cuda_dispatch.h>
+#include <ATen/ops/glu_cuda_dispatch.h>
+#include <ATen/ops/glu_backward_cuda_dispatch.h>
+#include <ATen/ops/glu_backward_jvp_cuda_dispatch.h>
+#include <ATen/ops/glu_jvp_cuda_dispatch.h>
+#include <ATen/ops/grid_sampler_2d_cuda_dispatch.h>
+#include <ATen/ops/grid_sampler_2d_backward_cuda_dispatch.h>
+#include <ATen/ops/grid_sampler_3d_cuda_dispatch.h>
+#include <ATen/ops/grid_sampler_3d_backward_cuda_dispatch.h>
+#include <ATen/ops/gt_cuda_dispatch.h>
+#include <ATen/ops/hardshrink_cuda_dispatch.h>
+#include <ATen/ops/hardshrink_backward_cuda_dispatch.h>
+#include <ATen/ops/hardsigmoid_cuda_dispatch.h>
+#include <ATen/ops/hardsigmoid_backward_cuda_dispatch.h>
+#include <ATen/ops/hardswish_cuda_dispatch.h>
+#include <ATen/ops/hardswish_backward_cuda_dispatch.h>
+#include <ATen/ops/hardtanh_cuda_dispatch.h>
+#include <ATen/ops/hardtanh_backward_cuda_dispatch.h>
+#include <ATen/ops/heaviside_cuda_dispatch.h>
+#include <ATen/ops/histc_cuda_dispatch.h>
+#include <ATen/ops/huber_loss_cuda_dispatch.h>
+#include <ATen/ops/huber_loss_backward_cuda_dispatch.h>
+#include <ATen/ops/hypot_cuda_dispatch.h>
+#include <ATen/ops/i0_cuda_dispatch.h>
+#include <ATen/ops/igamma_cuda_dispatch.h>
+#include <ATen/ops/igammac_cuda_dispatch.h>
+#include <ATen/ops/im2col_cuda_dispatch.h>
+#include <ATen/ops/index_cuda_dispatch.h>
+#include <ATen/ops/index_add_cuda_dispatch.h>
+#include <ATen/ops/index_copy_cuda_dispatch.h>
+#include <ATen/ops/index_fill_cuda_dispatch.h>
+#include <ATen/ops/index_reduce_cuda_dispatch.h>
+#include <ATen/ops/index_select_cuda_dispatch.h>
+#include <ATen/ops/is_pinned_cuda_dispatch.h>
+#include <ATen/ops/is_set_to_cuda_dispatch.h>
+#include <ATen/ops/isin_cuda_dispatch.h>
+#include <ATen/ops/isnan_cuda_dispatch.h>
+#include <ATen/ops/isneginf_cuda_dispatch.h>
+#include <ATen/ops/isposinf_cuda_dispatch.h>
+#include <ATen/ops/kthvalue_cuda_dispatch.h>
+#include <ATen/ops/lcm_cuda_dispatch.h>
+#include <ATen/ops/le_cuda_dispatch.h>
+#include <ATen/ops/leaky_relu_cuda_dispatch.h>
+#include <ATen/ops/leaky_relu_backward_cuda_dispatch.h>
+#include <ATen/ops/lerp_cuda_dispatch.h>
+#include <ATen/ops/lgamma_cuda_dispatch.h>
+#include <ATen/ops/linalg_cholesky_ex_cuda_dispatch.h>
+#include <ATen/ops/linalg_cross_cuda_dispatch.h>
+#include <ATen/ops/linalg_eig_cuda_dispatch.h>
+#include <ATen/ops/linalg_eigvals_cuda_dispatch.h>
+#include <ATen/ops/linalg_householder_product_cuda_dispatch.h>
+#include <ATen/ops/linalg_inv_ex_cuda_dispatch.h>
+#include <ATen/ops/linalg_ldl_factor_ex_cuda_dispatch.h>
+#include <ATen/ops/linalg_ldl_solve_cuda_dispatch.h>
+#include <ATen/ops/linalg_lstsq_cuda_dispatch.h>
+#include <ATen/ops/linalg_lu_cuda_dispatch.h>
+#include <ATen/ops/linalg_lu_factor_ex_cuda_dispatch.h>
+#include <ATen/ops/linalg_lu_solve_cuda_dispatch.h>
+#include <ATen/ops/linalg_matrix_exp_cuda_dispatch.h>
+#include <ATen/ops/linalg_qr_cuda_dispatch.h>
+#include <ATen/ops/linalg_solve_triangular_cuda_dispatch.h>
+#include <ATen/ops/linalg_vector_norm_cuda_dispatch.h>
+#include <ATen/ops/linspace_cuda_dispatch.h>
+#include <ATen/ops/log_cuda_dispatch.h>
+#include <ATen/ops/log10_cuda_dispatch.h>
+#include <ATen/ops/log1p_cuda_dispatch.h>
+#include <ATen/ops/log2_cuda_dispatch.h>
+#include <ATen/ops/log_normal_cuda_dispatch.h>
+#include <ATen/ops/log_sigmoid_backward_cuda_dispatch.h>
+#include <ATen/ops/log_sigmoid_forward_cuda_dispatch.h>
+#include <ATen/ops/logaddexp_cuda_dispatch.h>
+#include <ATen/ops/logaddexp2_cuda_dispatch.h>
+#include <ATen/ops/logical_and_cuda_dispatch.h>
+#include <ATen/ops/logical_not_cuda_dispatch.h>
+#include <ATen/ops/logical_or_cuda_dispatch.h>
+#include <ATen/ops/logical_xor_cuda_dispatch.h>
+#include <ATen/ops/logit_cuda_dispatch.h>
+#include <ATen/ops/logit_backward_cuda_dispatch.h>
+#include <ATen/ops/logspace_cuda_dispatch.h>
+#include <ATen/ops/lshift_cuda_dispatch.h>
+#include <ATen/ops/lt_cuda_dispatch.h>
+#include <ATen/ops/lu_unpack_cuda_dispatch.h>
+#include <ATen/ops/masked_fill_cuda_dispatch.h>
+#include <ATen/ops/masked_scatter_cuda_dispatch.h>
+#include <ATen/ops/masked_select_cuda_dispatch.h>
+#include <ATen/ops/max_cuda_dispatch.h>
+#include <ATen/ops/max_pool2d_with_indices_cuda_dispatch.h>
+#include <ATen/ops/max_pool2d_with_indices_backward_cuda_dispatch.h>
+#include <ATen/ops/max_pool3d_with_indices_cuda_dispatch.h>
+#include <ATen/ops/max_pool3d_with_indices_backward_cuda_dispatch.h>
+#include <ATen/ops/max_unpool2d_cuda_dispatch.h>
+#include <ATen/ops/max_unpool3d_cuda_dispatch.h>
+#include <ATen/ops/maximum_cuda_dispatch.h>
+#include <ATen/ops/mean_cuda_dispatch.h>
+#include <ATen/ops/median_cuda_dispatch.h>
+#include <ATen/ops/min_cuda_dispatch.h>
+#include <ATen/ops/minimum_cuda_dispatch.h>
+#include <ATen/ops/miopen_batch_norm_cuda_dispatch.h>
+#include <ATen/ops/miopen_batch_norm_backward_cuda_dispatch.h>
+#include <ATen/ops/miopen_convolution_cuda_dispatch.h>
+#include <ATen/ops/miopen_convolution_add_relu_cuda_dispatch.h>
+#include <ATen/ops/miopen_convolution_relu_cuda_dispatch.h>
+#include <ATen/ops/miopen_convolution_transpose_cuda_dispatch.h>
+#include <ATen/ops/miopen_depthwise_convolution_cuda_dispatch.h>
+#include <ATen/ops/miopen_rnn_cuda_dispatch.h>
+#include <ATen/ops/miopen_rnn_backward_cuda_dispatch.h>
+#include <ATen/ops/mish_cuda_dispatch.h>
+#include <ATen/ops/mish_backward_cuda_dispatch.h>
+#include <ATen/ops/mm_cuda_dispatch.h>
+#include <ATen/ops/mode_cuda_dispatch.h>
+#include <ATen/ops/mse_loss_cuda_dispatch.h>
+#include <ATen/ops/mse_loss_backward_cuda_dispatch.h>
+#include <ATen/ops/mul_cuda_dispatch.h>
+#include <ATen/ops/multi_margin_loss_cuda_dispatch.h>
+#include <ATen/ops/multi_margin_loss_backward_cuda_dispatch.h>
+#include <ATen/ops/multilabel_margin_loss_backward_cuda_dispatch.h>
+#include <ATen/ops/multilabel_margin_loss_forward_cuda_dispatch.h>
+#include <ATen/ops/multinomial_cuda_dispatch.h>
+#include <ATen/ops/mvlgamma_cuda_dispatch.h>
+#include <ATen/ops/nan_to_num_cuda_dispatch.h>
+#include <ATen/ops/nanmedian_cuda_dispatch.h>
+#include <ATen/ops/nansum_cuda_dispatch.h>
+#include <ATen/ops/native_batch_norm_cuda_dispatch.h>
+#include <ATen/ops/native_batch_norm_backward_cuda_dispatch.h>
+#include <ATen/ops/native_dropout_cuda_dispatch.h>
+#include <ATen/ops/native_dropout_backward_cuda_dispatch.h>
+#include <ATen/ops/native_group_norm_cuda_dispatch.h>
+#include <ATen/ops/native_group_norm_backward_cuda_dispatch.h>
+#include <ATen/ops/native_layer_norm_cuda_dispatch.h>
+#include <ATen/ops/native_layer_norm_backward_cuda_dispatch.h>
+#include <ATen/ops/ne_cuda_dispatch.h>
+#include <ATen/ops/neg_cuda_dispatch.h>
+#include <ATen/ops/nextafter_cuda_dispatch.h>
+#include <ATen/ops/nll_loss2d_backward_cuda_dispatch.h>
+#include <ATen/ops/nll_loss2d_forward_cuda_dispatch.h>
+#include <ATen/ops/nll_loss_backward_cuda_dispatch.h>
+#include <ATen/ops/nll_loss_forward_cuda_dispatch.h>
+#include <ATen/ops/nonzero_cuda_dispatch.h>
+#include <ATen/ops/norm_cuda_dispatch.h>
+#include <ATen/ops/normal_cuda_dispatch.h>
+#include <ATen/ops/ormqr_cuda_dispatch.h>
+#include <ATen/ops/poisson_cuda_dispatch.h>
+#include <ATen/ops/polar_cuda_dispatch.h>
+#include <ATen/ops/polygamma_cuda_dispatch.h>
+#include <ATen/ops/pow_cuda_dispatch.h>
+#include <ATen/ops/prod_cuda_dispatch.h>
+#include <ATen/ops/put_cuda_dispatch.h>
+#include <ATen/ops/quantize_per_channel_cuda_dispatch.h>
+#include <ATen/ops/quantize_per_tensor_cuda_dispatch.h>
+#include <ATen/ops/quantize_per_tensor_dynamic_cuda_dispatch.h>
+#include <ATen/ops/random_cuda_dispatch.h>
+#include <ATen/ops/randperm_cuda_dispatch.h>
+#include <ATen/ops/range_cuda_dispatch.h>
+#include <ATen/ops/reciprocal_cuda_dispatch.h>
+#include <ATen/ops/record_stream_cuda_dispatch.h>
+#include <ATen/ops/reflection_pad1d_cuda_dispatch.h>
+#include <ATen/ops/reflection_pad1d_backward_cuda_dispatch.h>
+#include <ATen/ops/reflection_pad2d_cuda_dispatch.h>
+#include <ATen/ops/reflection_pad2d_backward_cuda_dispatch.h>
+#include <ATen/ops/reflection_pad3d_cuda_dispatch.h>
+#include <ATen/ops/reflection_pad3d_backward_cuda_dispatch.h>
+#include <ATen/ops/relu_cuda_dispatch.h>
+#include <ATen/ops/remainder_cuda_dispatch.h>
+#include <ATen/ops/renorm_cuda_dispatch.h>
+#include <ATen/ops/repeat_interleave_cuda_dispatch.h>
+#include <ATen/ops/replication_pad1d_cuda_dispatch.h>
+#include <ATen/ops/replication_pad1d_backward_cuda_dispatch.h>
+#include <ATen/ops/replication_pad2d_cuda_dispatch.h>
+#include <ATen/ops/replication_pad2d_backward_cuda_dispatch.h>
+#include <ATen/ops/replication_pad3d_cuda_dispatch.h>
+#include <ATen/ops/replication_pad3d_backward_cuda_dispatch.h>
+#include <ATen/ops/resize_cuda_dispatch.h>
+#include <ATen/ops/roll_cuda_dispatch.h>
+#include <ATen/ops/round_cuda_dispatch.h>
+#include <ATen/ops/rrelu_with_noise_cuda_dispatch.h>
+#include <ATen/ops/rshift_cuda_dispatch.h>
+#include <ATen/ops/rsqrt_cuda_dispatch.h>
+#include <ATen/ops/rsub_cuda_dispatch.h>
+#include <ATen/ops/scatter_cuda_dispatch.h>
+#include <ATen/ops/scatter_add_cuda_dispatch.h>
+#include <ATen/ops/scatter_reduce_cuda_dispatch.h>
+#include <ATen/ops/searchsorted_cuda_dispatch.h>
+#include <ATen/ops/segment_reduce_cuda_dispatch.h>
+#include <ATen/ops/set_cuda_dispatch.h>
+#include <ATen/ops/sgn_cuda_dispatch.h>
+#include <ATen/ops/sigmoid_cuda_dispatch.h>
+#include <ATen/ops/sigmoid_backward_cuda_dispatch.h>
+#include <ATen/ops/sign_cuda_dispatch.h>
+#include <ATen/ops/signbit_cuda_dispatch.h>
+#include <ATen/ops/silu_cuda_dispatch.h>
+#include <ATen/ops/silu_backward_cuda_dispatch.h>
+#include <ATen/ops/sin_cuda_dispatch.h>
+#include <ATen/ops/sinc_cuda_dispatch.h>
+#include <ATen/ops/sinh_cuda_dispatch.h>
+#include <ATen/ops/slow_conv_dilated2d_cuda_dispatch.h>
+#include <ATen/ops/slow_conv_dilated3d_cuda_dispatch.h>
+#include <ATen/ops/slow_conv_transpose2d_cuda_dispatch.h>
+#include <ATen/ops/slow_conv_transpose3d_cuda_dispatch.h>
+#include <ATen/ops/smooth_l1_loss_cuda_dispatch.h>
+#include <ATen/ops/smooth_l1_loss_backward_cuda_dispatch.h>
+#include <ATen/ops/softplus_cuda_dispatch.h>
+#include <ATen/ops/softplus_backward_cuda_dispatch.h>
+#include <ATen/ops/softshrink_cuda_dispatch.h>
+#include <ATen/ops/softshrink_backward_cuda_dispatch.h>
+#include <ATen/ops/sort_cuda_dispatch.h>
+#include <ATen/ops/sparse_dim_cuda_dispatch.h>
+#include <ATen/ops/special_airy_ai_cuda_dispatch.h>
+#include <ATen/ops/special_bessel_j0_cuda_dispatch.h>
+#include <ATen/ops/special_bessel_j1_cuda_dispatch.h>
+#include <ATen/ops/special_bessel_y0_cuda_dispatch.h>
+#include <ATen/ops/special_bessel_y1_cuda_dispatch.h>
+#include <ATen/ops/special_chebyshev_polynomial_t_cuda_dispatch.h>
+#include <ATen/ops/special_chebyshev_polynomial_u_cuda_dispatch.h>
+#include <ATen/ops/special_chebyshev_polynomial_v_cuda_dispatch.h>
+#include <ATen/ops/special_chebyshev_polynomial_w_cuda_dispatch.h>
+#include <ATen/ops/special_entr_cuda_dispatch.h>
+#include <ATen/ops/special_erfcx_cuda_dispatch.h>
+#include <ATen/ops/special_hermite_polynomial_h_cuda_dispatch.h>
+#include <ATen/ops/special_hermite_polynomial_he_cuda_dispatch.h>
+#include <ATen/ops/special_i0e_cuda_dispatch.h>
+#include <ATen/ops/special_i1_cuda_dispatch.h>
+#include <ATen/ops/special_i1e_cuda_dispatch.h>
+#include <ATen/ops/special_laguerre_polynomial_l_cuda_dispatch.h>
+#include <ATen/ops/special_legendre_polynomial_p_cuda_dispatch.h>
+#include <ATen/ops/special_log_ndtr_cuda_dispatch.h>
+#include <ATen/ops/special_modified_bessel_i0_cuda_dispatch.h>
+#include <ATen/ops/special_modified_bessel_i1_cuda_dispatch.h>
+#include <ATen/ops/special_modified_bessel_k0_cuda_dispatch.h>
+#include <ATen/ops/special_modified_bessel_k1_cuda_dispatch.h>
+#include <ATen/ops/special_ndtri_cuda_dispatch.h>
+#include <ATen/ops/special_scaled_modified_bessel_k0_cuda_dispatch.h>
+#include <ATen/ops/special_scaled_modified_bessel_k1_cuda_dispatch.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_t_cuda_dispatch.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_u_cuda_dispatch.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_v_cuda_dispatch.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_w_cuda_dispatch.h>
+#include <ATen/ops/special_spherical_bessel_j0_cuda_dispatch.h>
+#include <ATen/ops/special_xlog1py_cuda_dispatch.h>
+#include <ATen/ops/special_zeta_cuda_dispatch.h>
+#include <ATen/ops/split_with_sizes_copy_cuda_dispatch.h>
+#include <ATen/ops/sqrt_cuda_dispatch.h>
+#include <ATen/ops/sspaddmm_cuda_dispatch.h>
+#include <ATen/ops/std_cuda_dispatch.h>
+#include <ATen/ops/std_mean_cuda_dispatch.h>
+#include <ATen/ops/sub_cuda_dispatch.h>
+#include <ATen/ops/sum_cuda_dispatch.h>
+#include <ATen/ops/take_cuda_dispatch.h>
+#include <ATen/ops/tan_cuda_dispatch.h>
+#include <ATen/ops/tanh_cuda_dispatch.h>
+#include <ATen/ops/tanh_backward_cuda_dispatch.h>
+#include <ATen/ops/threshold_cuda_dispatch.h>
+#include <ATen/ops/threshold_backward_cuda_dispatch.h>
+#include <ATen/ops/topk_cuda_dispatch.h>
+#include <ATen/ops/trace_cuda_dispatch.h>
+#include <ATen/ops/triangular_solve_cuda_dispatch.h>
+#include <ATen/ops/tril_cuda_dispatch.h>
+#include <ATen/ops/tril_indices_cuda_dispatch.h>
+#include <ATen/ops/triu_cuda_dispatch.h>
+#include <ATen/ops/triu_indices_cuda_dispatch.h>
+#include <ATen/ops/trunc_cuda_dispatch.h>
+#include <ATen/ops/unfold_cuda_dispatch.h>
+#include <ATen/ops/unfold_backward_cuda_dispatch.h>
+#include <ATen/ops/uniform_cuda_dispatch.h>
+#include <ATen/ops/unique_consecutive_cuda_dispatch.h>
+#include <ATen/ops/unique_dim_cuda_dispatch.h>
+#include <ATen/ops/unique_dim_consecutive_cuda_dispatch.h>
+#include <ATen/ops/upsample_bicubic2d_cuda_dispatch.h>
+#include <ATen/ops/upsample_bicubic2d_backward_cuda_dispatch.h>
+#include <ATen/ops/upsample_bilinear2d_cuda_dispatch.h>
+#include <ATen/ops/upsample_bilinear2d_backward_cuda_dispatch.h>
+#include <ATen/ops/upsample_linear1d_cuda_dispatch.h>
+#include <ATen/ops/upsample_linear1d_backward_cuda_dispatch.h>
+#include <ATen/ops/upsample_nearest1d_cuda_dispatch.h>
+#include <ATen/ops/upsample_nearest1d_backward_cuda_dispatch.h>
+#include <ATen/ops/upsample_nearest2d_cuda_dispatch.h>
+#include <ATen/ops/upsample_nearest2d_backward_cuda_dispatch.h>
+#include <ATen/ops/upsample_nearest3d_cuda_dispatch.h>
+#include <ATen/ops/upsample_nearest3d_backward_cuda_dispatch.h>
+#include <ATen/ops/upsample_trilinear3d_cuda_dispatch.h>
+#include <ATen/ops/upsample_trilinear3d_backward_cuda_dispatch.h>
+#include <ATen/ops/var_cuda_dispatch.h>
+#include <ATen/ops/var_mean_cuda_dispatch.h>
+#include <ATen/ops/vdot_cuda_dispatch.h>
+#include <ATen/ops/view_cuda_dispatch.h>
+#include <ATen/ops/view_as_complex_cuda_dispatch.h>
+#include <ATen/ops/view_as_real_cuda_dispatch.h>
+#include <ATen/ops/where_cuda_dispatch.h>
+#include <ATen/ops/xlogy_cuda_dispatch.h>
+#include <ATen/ops/zero_cuda_dispatch.h>
+
+
+

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/CachedTensorUtils.h

@@ -0,0 +1,24 @@
+#pragma once
+
+#include <ATen/ATen.h>
+
+namespace at::caching {
+
+// Some systems (just cudagraphs currently) will persist a static tensor output
+// whose TensorImpl does not change across iterations. For these tensors caching
+// dtype conversions is invalid. Additionally, there will be an extra reference
+// count to these cached tensors that would prevent buffer inplacing and other
+// checks on tensor uniqueness. If we are not using these systems the enabled
+// flag will be false and we will avoid the hash lookup.
+
+TORCH_API bool is_cached_tensor(const at::Tensor& t);
+TORCH_API void add_cached_tensor(const at::Tensor& t);
+TORCH_API void remove_cached_tensor(const at::Tensor& t);
+TORCH_API void set_cached_tensors_enabled(bool enable);
+
+// For gradient buffer stealing we will adjust the use count of tensors
+// which are persisted by cudagraphs, just as we need to adjust reference
+// count of tensors with hooks.
+TORCH_API size_t adjusted_use_count(const at::Tensor& t);
+
+} // namespace at::caching

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/CompositeExplicitAutogradFunctions.h

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

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/CompositeExplicitAutogradNonFunctionalFunctions.h

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

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/CompositeExplicitAutogradNonFunctionalFunctions_inl.h

@@ -0,0 +1,323 @@
+#pragma once
+// @generated by torchgen/gen.py from DispatchKeyFunctions_inl.h
+
+// NB: The implementing C++ file is RegisterDispatchKey.cpp
+
+// The only #includes we need are for custom classes that have defaults in the C++ API
+#include <c10/core/MemoryFormat.h>
+#include <c10/core/Scalar.h>
+#include <ATen/core/Reduction.h>
+
+#if defined(AT_PER_OPERATOR_HEADERS) && defined(TORCH_ASSERT_ONLY_METHOD_OPERATORS)
+#error This change adds a dependency on all pytorch operators, meaning the     \
+  file will need to be re-compiled every time an operator is changed or added. \
+  Consider including a specific operator from                                  \
+  <ATen/ops/{my_operator}_compositeexplicitautogradnonfunctional_dispatch.h>.                   \
+  See NOTE [TORCH_ASSERT_ONLY_METHOD_OPERATORS].
+#endif
+
+#include <ATen/ops/_addmm_activation_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_conj_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_convert_indices_from_coo_to_csr_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_convert_indices_from_csr_to_coo_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_fw_primal_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_indices_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_linalg_det_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_linalg_eigh_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_linalg_slogdet_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_linalg_solve_ex_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_linalg_svd_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_log_softmax_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_log_softmax_backward_data_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_make_dual_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_neg_view_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_nested_get_values_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_nested_view_from_buffer_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_nested_view_from_jagged_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_reshape_alias_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_softmax_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_softmax_backward_data_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_sparse_broadcast_to_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_test_autograd_multiple_dispatch_view_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_trilinear_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_upsample_bicubic2d_aa_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_upsample_bicubic2d_aa_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_upsample_bilinear2d_aa_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_upsample_bilinear2d_aa_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_upsample_nearest_exact1d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_upsample_nearest_exact1d_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_upsample_nearest_exact2d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_upsample_nearest_exact2d_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_upsample_nearest_exact3d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_upsample_nearest_exact3d_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/_values_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/acos_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/acosh_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/adaptive_max_pool2d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/adaptive_max_pool2d_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/adaptive_max_pool3d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/adaptive_max_pool3d_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/add_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/addcdiv_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/addcmul_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/addmm_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/addmv_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/alias_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/all_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/amax_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/amin_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/aminmax_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/any_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/argmax_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/argmin_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/as_strided_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/as_strided_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/as_strided_scatter_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/asin_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/asinh_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/atan_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/atan2_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/atanh_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/avg_pool2d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/avg_pool2d_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/avg_pool3d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/avg_pool3d_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/baddbmm_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/bernoulli_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/bitwise_and_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/bitwise_left_shift_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/bitwise_not_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/bitwise_or_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/bitwise_right_shift_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/bitwise_xor_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/bmm_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/cat_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/ccol_indices_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/ceil_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/clamp_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/clamp_max_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/clamp_min_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/col_indices_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/copysign_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/cos_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/cosh_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/crow_indices_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/cumprod_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/cumsum_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/detach_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/diag_embed_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/diagonal_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/diagonal_scatter_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/digamma_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/div_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/elu_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/elu_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/eq_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/erf_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/erfc_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/erfinv_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/exp_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/exp2_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/expand_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/expm1_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/floor_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/fmax_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/fmin_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/fmod_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/frac_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/fractional_max_pool2d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/fractional_max_pool2d_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/fractional_max_pool3d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/gather_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/gcd_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/ge_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/gelu_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/gelu_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/glu_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/gt_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/hardshrink_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/hardshrink_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/hardsigmoid_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/hardsigmoid_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/heaviside_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/hypot_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/i0_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/igamma_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/igammac_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/index_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/index_add_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/index_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/index_reduce_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/indices_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/isin_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/isneginf_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/isposinf_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/lcm_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/le_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/leaky_relu_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/leaky_relu_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/lerp_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/lgamma_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/lift_fresh_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/linalg_cholesky_ex_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/linalg_cross_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/linalg_inv_ex_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/linalg_ldl_factor_ex_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/linalg_ldl_solve_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/linalg_lu_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/linalg_lu_factor_ex_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/linalg_lu_solve_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/linalg_pinv_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/linalg_qr_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/linalg_vector_norm_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/log_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/log10_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/log1p_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/log2_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/logaddexp_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/logaddexp2_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/logit_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/logsumexp_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/lt_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/lu_unpack_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/max_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/max_pool2d_with_indices_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/max_pool2d_with_indices_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/maximum_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/mean_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/min_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/minimum_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/mish_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/mm_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/mse_loss_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/mul_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/narrow_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/ne_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/neg_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/new_empty_strided_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/nextafter_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/nll_loss_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/nll_loss_forward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/norm_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/permute_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/pixel_shuffle_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/pixel_unshuffle_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/polygamma_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/pow_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/prod_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/reciprocal_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/reflection_pad1d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/reflection_pad1d_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/reflection_pad3d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/reflection_pad3d_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/remainder_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/renorm_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/replication_pad1d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/replication_pad1d_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/replication_pad2d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/replication_pad3d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/round_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/row_indices_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/rsqrt_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/scatter_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/scatter_add_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/scatter_reduce_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/select_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/select_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/select_scatter_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/sgn_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/sigmoid_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/sigmoid_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/sign_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/signbit_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/silu_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/silu_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/sin_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/sinc_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/sinh_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/slice_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/slice_scatter_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/slow_conv_transpose2d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/smooth_l1_loss_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/softplus_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/softplus_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/softshrink_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/softshrink_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/sort_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_airy_ai_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_bessel_j0_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_bessel_j1_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_bessel_y0_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_bessel_y1_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_chebyshev_polynomial_t_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_chebyshev_polynomial_u_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_chebyshev_polynomial_v_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_chebyshev_polynomial_w_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_entr_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_erfcx_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_hermite_polynomial_h_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_hermite_polynomial_he_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_i0e_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_i1_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_i1e_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_laguerre_polynomial_l_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_legendre_polynomial_p_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_log_ndtr_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_modified_bessel_i0_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_modified_bessel_i1_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_modified_bessel_k0_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_modified_bessel_k1_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_ndtri_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_scaled_modified_bessel_k0_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_scaled_modified_bessel_k1_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_t_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_u_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_v_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_w_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_spherical_bessel_j0_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_xlog1py_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/special_zeta_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/split_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/split_with_sizes_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/sqrt_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/squeeze_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/sub_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/sum_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/t_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/tan_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/tanh_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/tanh_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/threshold_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/threshold_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/topk_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/transpose_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/triangular_solve_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/tril_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/triu_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/trunc_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/unbind_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/unfold_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/unsqueeze_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/upsample_bicubic2d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/upsample_bicubic2d_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/upsample_bilinear2d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/upsample_bilinear2d_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/upsample_linear1d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/upsample_linear1d_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/upsample_nearest1d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/upsample_nearest1d_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/upsample_nearest2d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/upsample_nearest2d_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/upsample_nearest3d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/upsample_nearest3d_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/upsample_trilinear3d_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/upsample_trilinear3d_backward_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/values_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/view_as_complex_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/view_as_real_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/view_copy_compositeexplicitautogradnonfunctional_dispatch.h>
+#include <ATen/ops/xlogy_compositeexplicitautogradnonfunctional_dispatch.h>
+
+
+

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/CompositeImplicitAutogradFunctions_inl.h

@@ -0,0 +1,500 @@
+#pragma once
+// @generated by torchgen/gen.py from DispatchKeyFunctions_inl.h
+
+// NB: The implementing C++ file is RegisterDispatchKey.cpp
+
+// The only #includes we need are for custom classes that have defaults in the C++ API
+#include <c10/core/MemoryFormat.h>
+#include <c10/core/Scalar.h>
+#include <ATen/core/Reduction.h>
+
+#if defined(AT_PER_OPERATOR_HEADERS) && defined(TORCH_ASSERT_ONLY_METHOD_OPERATORS)
+#error This change adds a dependency on all pytorch operators, meaning the     \
+  file will need to be re-compiled every time an operator is changed or added. \
+  Consider including a specific operator from                                  \
+  <ATen/ops/{my_operator}_compositeimplicitautograd_dispatch.h>.                   \
+  See NOTE [TORCH_ASSERT_ONLY_METHOD_OPERATORS].
+#endif
+
+#include <ATen/ops/_add_batch_dim_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_assert_tensor_metadata_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_autocast_to_full_precision_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_autocast_to_reduced_precision_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_batch_norm_impl_index_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_batch_norm_impl_index_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_cast_Byte_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_cast_Char_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_cast_Double_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_cast_Float_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_cast_Half_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_cast_Int_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_cast_Long_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_cast_Short_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_choose_qparams_per_tensor_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_convolution_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_convolution_double_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_convolution_mode_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_cufft_clear_plan_cache_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_cufft_get_plan_cache_max_size_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_cufft_get_plan_cache_size_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_cufft_set_plan_cache_max_size_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_debug_has_internal_overlap_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_dim_arange_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_embedding_bag_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_embedding_bag_sparse_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_gather_sparse_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_grid_sampler_2d_cpu_fallback_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_has_compatible_shallow_copy_type_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_is_zerotensor_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_lu_with_info_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_nnpack_available_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_pack_padded_sequence_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_pad_circular_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_pad_enum_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_pad_packed_sequence_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_propagate_xla_data_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_remove_batch_dim_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_reshape_from_tensor_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_rowwise_prune_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_saturate_weight_to_fp16_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_scaled_dot_product_attention_math_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_shape_as_tensor_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_sobol_engine_draw_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_sobol_engine_ff_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_sobol_engine_initialize_state_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_sobol_engine_scramble_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_bsc_tensor_unsafe_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_bsr_tensor_unsafe_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_compressed_tensor_unsafe_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_coo_tensor_unsafe_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_csc_tensor_unsafe_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_csr_tensor_unsafe_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_log_softmax_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_mm_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_softmax_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_sparse_sum_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_test_ambiguous_defaults_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_test_autograd_multiple_dispatch_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_test_check_tensor_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_test_serialization_subcmul_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_test_string_default_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_thnn_differentiable_gru_cell_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_thnn_differentiable_lstm_cell_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_thnn_fused_lstm_cell_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_to_cpu_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_unpack_dual_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_upsample_bicubic2d_aa_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_upsample_bilinear2d_aa_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_upsample_nearest_exact1d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_upsample_nearest_exact2d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_upsample_nearest_exact3d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_use_cudnn_rnn_flatten_weight_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_validate_sparse_bsc_tensor_args_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_validate_sparse_bsr_tensor_args_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_validate_sparse_compressed_tensor_args_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_validate_sparse_coo_tensor_args_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_validate_sparse_csc_tensor_args_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_validate_sparse_csr_tensor_args_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_version_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_weight_norm_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/_weight_norm_differentiable_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/absolute_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/adaptive_avg_pool1d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/adaptive_avg_pool2d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/adaptive_avg_pool3d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/adaptive_max_pool1d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/adjoint_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/affine_grid_generator_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/align_as_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/align_tensors_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/align_to_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/all_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/alpha_dropout_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/and_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/any_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/arccos_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/arccosh_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/arcsin_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/arcsinh_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/arctan_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/arctan2_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/arctanh_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/argsort_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/argwhere_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/atleast_1d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/atleast_2d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/atleast_3d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/avg_pool1d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/batch_norm_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/bilinear_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/broadcast_tensors_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/broadcast_to_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/can_cast_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/cartesian_prod_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/cat_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/cdist_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/chain_matmul_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/chalf_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/choose_qparams_optimized_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/chunk_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/clip_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/coalesce_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/column_stack_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/combinations_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/concat_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/concatenate_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/conj_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/conj_physical_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/contiguous_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/conv1d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/conv2d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/conv3d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/conv_tbc_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/conv_transpose1d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/conv_transpose2d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/conv_transpose3d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/corrcoef_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/cosine_embedding_loss_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/cosine_similarity_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/cov_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/cross_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/cross_entropy_loss_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/ctc_loss_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/cudnn_is_acceptable_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/cummax_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/cummaxmin_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/cummin_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/cumprod_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/cumprod_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/cumsum_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/cumulative_trapezoid_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/data_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/det_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/diag_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/diagflat_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/diagonal_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/diff_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/divide_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/dropout_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/dsplit_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/dstack_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/einsum_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/embedding_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/embedding_bag_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/embedding_sparse_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/empty_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/expand_as_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fake_quantize_per_channel_affine_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fake_quantize_per_channel_affine_cachemask_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fake_quantize_per_tensor_affine_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fake_quantize_per_tensor_affine_cachemask_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fbgemm_linear_fp16_weight_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fbgemm_linear_fp16_weight_fp32_activation_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fbgemm_linear_int8_weight_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fbgemm_linear_int8_weight_fp32_activation_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fbgemm_linear_quantize_weight_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fbgemm_pack_gemm_matrix_fp16_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fbgemm_pack_quantized_matrix_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/feature_alpha_dropout_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/feature_dropout_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fft_fft_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fft_fft2_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fft_fftn_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fft_fftshift_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fft_hfft_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fft_hfft2_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fft_hfftn_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fft_ifft_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fft_ifft2_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fft_ifftn_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fft_ifftshift_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fft_ihfft_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fft_ihfft2_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fft_ihfftn_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fft_irfft_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fft_irfft2_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fft_irfftn_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fft_rfft_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fft_rfft2_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fft_rfftn_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fill_diagonal_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fix_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/flatten_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/flatten_dense_tensors_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fliplr_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/flipud_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/float_power_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/frobenius_norm_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/fused_moving_avg_obs_fake_quant_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/gather_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/gather_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/ger_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/gradient_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/greater_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/greater_equal_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/grid_sampler_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/group_norm_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/gru_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/gru_cell_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/hinge_embedding_loss_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/histogramdd_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/hsplit_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/hstack_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/imag_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/index_add_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/index_copy_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/index_fill_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/index_select_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/index_select_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/infinitely_differentiable_gelu_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/inner_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/instance_norm_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/inverse_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/is_complex_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/is_conj_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/is_distributed_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/is_floating_point_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/is_inference_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/is_leaf_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/is_neg_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/is_nonzero_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/is_signed_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/is_vulkan_available_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/isclose_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/isfinite_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/isreal_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/istft_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/item_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/kl_div_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/kron_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/kthvalue_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/l1_loss_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/layer_norm_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/ldexp_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/less_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/less_equal_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_cholesky_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_cond_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_det_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_diagonal_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_eigh_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_eigvals_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_eigvalsh_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_inv_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_ldl_factor_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_lu_factor_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_matmul_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_matrix_norm_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_matrix_power_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_matrix_rank_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_multi_dot_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_norm_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_pinv_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_slogdet_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_solve_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_solve_ex_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_svd_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_svdvals_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_tensorinv_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_tensorsolve_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_vander_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linalg_vecdot_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/linear_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/log_sigmoid_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/log_softmax_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/logcumsumexp_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/logdet_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/logsumexp_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/lstm_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/lstm_cell_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/lu_solve_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/mH_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/mT_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/margin_ranking_loss_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/masked_select_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/matmul_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/matrix_H_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/matrix_exp_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/matrix_exp_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/matrix_power_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/max_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/max_pool1d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/max_pool1d_with_indices_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/max_pool2d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/max_pool3d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/mean_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/median_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/meshgrid_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/min_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/mish_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/mode_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/moveaxis_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/movedim_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/msort_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/multilabel_margin_loss_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/multiply_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/nanmean_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/nanmedian_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/nanquantile_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/narrow_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/native_channel_shuffle_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/negative_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/nested_to_padded_tensor_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/nll_loss_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/nll_loss2d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/nll_loss_nd_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/nonzero_numpy_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/norm_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/norm_except_dim_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/not_equal_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/nuclear_norm_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/numpy_T_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/one_hot_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/or_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/orgqr_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/outer_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/output_nr_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/pad_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/pad_sequence_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/pairwise_distance_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/pdist_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/pin_memory_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/pinverse_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/poisson_nll_loss_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/positive_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/prelu_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/prod_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/promote_types_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/qr_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/quantile_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/quantized_gru_cell_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/quantized_lstm_cell_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/quantized_rnn_relu_cell_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/quantized_rnn_tanh_cell_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/rand_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/randn_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/ravel_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/real_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/refine_names_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/relu6_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/rename_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/repeat_interleave_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/requires_grad_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/reshape_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/reshape_as_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/resolve_conj_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/resolve_neg_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/result_type_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/retain_grad_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/retains_grad_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/rnn_relu_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/rnn_relu_cell_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/rnn_tanh_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/rnn_tanh_cell_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/row_stack_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/rrelu_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/scaled_dot_product_attention_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/scatter_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/scatter_add_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/select_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/selu_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/set_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/set_data_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/silu_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/size_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/slogdet_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/slow_conv3d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/smm_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/softmax_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/sort_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/sparse_bsc_tensor_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/sparse_bsr_tensor_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/sparse_coo_tensor_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/sparse_csc_tensor_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/sparse_csr_tensor_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_digamma_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_erf_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_erfc_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_erfinv_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_exp2_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_expit_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_expm1_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_gammainc_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_gammaincc_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_gammaln_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_i0_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_log1p_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_log_softmax_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_logit_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_logsumexp_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_multigammaln_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_ndtr_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_polygamma_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_psi_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_round_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_sinc_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_softmax_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/special_xlogy_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/split_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/square_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/squeeze_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/sspaddmm_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/std_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/std_mean_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/stft_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/stride_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/subtract_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/sum_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/sum_to_size_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/svd_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/swapaxes_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/swapdims_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/sym_numel_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/sym_size_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/sym_storage_offset_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/sym_stride_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/take_along_dim_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/tensor_split_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/tensordot_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/thnn_conv2d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/tile_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/to_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/to_dense_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/to_dense_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/to_mkldnn_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/to_sparse_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/to_sparse_bsc_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/to_sparse_bsr_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/to_sparse_csc_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/to_sparse_csr_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/trace_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/transpose_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/trapezoid_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/trapz_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/triplet_margin_loss_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/true_divide_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/type_as_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/unbind_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/unflatten_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/unflatten_dense_tensors_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/unsafe_chunk_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/upsample_bicubic2d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/upsample_bilinear2d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/upsample_linear1d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/upsample_nearest1d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/upsample_nearest2d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/upsample_nearest3d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/upsample_trilinear3d_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/value_selecting_reduction_backward_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/vander_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/var_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/var_mean_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/view_as_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/vsplit_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/vstack_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/where_compositeimplicitautograd_dispatch.h>
+#include <ATen/ops/xor_compositeimplicitautograd_dispatch.h>
+
+
+

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

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

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/DLConvertor.h

@@ -0,0 +1,25 @@
+#pragma once
+
+#include <ATen/ATen.h>
+#include <ATen/Tensor.h>
+#include <ATen/dlpack.h>
+
+// this convertor will:
+// 1) take a Tensor object and wrap it in the DLPack tensor
+// 2) take a dlpack tensor and convert it to the ATen Tensor
+
+namespace at {
+
+TORCH_API ScalarType toScalarType(const DLDataType& dtype);
+TORCH_API DLManagedTensor* toDLPack(const Tensor& src);
+TORCH_API Tensor fromDLPack(DLManagedTensor* src);
+C10_DEPRECATED_MESSAGE("Please migrate to a non-const variant")
+inline Tensor fromDLPack(const DLManagedTensor* src) {
+  return fromDLPack(const_cast<DLManagedTensor*>(src));
+}
+TORCH_API Tensor
+fromDLPack(DLManagedTensor* src, std::function<void(void*)> deleter);
+TORCH_API DLDataType getDLDataType(const Tensor& t);
+TORCH_API DLDevice getDLContext(const Tensor& tensor, const int64_t& device_id);
+
+} // namespace at

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

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

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/DeviceGuard.h

@@ -0,0 +1,41 @@
+#pragma once
+
+#include <ATen/core/IListRef.h>
+#include <ATen/core/Tensor.h>
+#include <c10/core/DeviceGuard.h>
+#include <c10/core/ScalarType.h> // TensorList whyyyyy
+
+namespace at {
+
+// Are you here because you're wondering why DeviceGuard(tensor) no
+// longer works?  For code organization reasons, we have temporarily(?)
+// removed this constructor from DeviceGuard.  The new way to
+// spell it is:
+//
+//    OptionalDeviceGuard guard(device_of(tensor));
+
+/// Return the Device of a Tensor, if the Tensor is defined.
+inline c10::optional<Device> device_of(const Tensor& t) {
+  if (t.defined()) {
+    return c10::make_optional(t.device());
+  } else {
+    return c10::nullopt;
+  }
+}
+
+inline c10::optional<Device> device_of(const c10::optional<Tensor>& t) {
+  return t.has_value() ? device_of(t.value()) : c10::nullopt;
+}
+
+/// Return the Device of a TensorList, if the list is non-empty and
+/// the first Tensor is defined.  (This function implicitly assumes
+/// that all tensors in the list have the same device.)
+inline c10::optional<Device> device_of(ITensorListRef t) {
+  if (!t.empty()) {
+    return device_of(t.front());
+  } else {
+    return c10::nullopt;
+  }
+}
+
+} // namespace at

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/ExpandUtils.h

@@ -0,0 +1,527 @@
+#pragma once
+
+#ifndef AT_PER_OPERATOR_HEADERS
+#include <ATen/Functions.h>
+#else
+#include <ATen/ops/view.h>
+#include <ATen/ops/view_copy.h>
+#endif
+
+#include <ATen/Tensor.h>
+#include <ATen/core/DimVector.h>
+#include <c10/util/Exception.h>
+#include <c10/util/MaybeOwned.h>
+#include <c10/util/irange.h>
+
+#include <functional>
+#include <sstream>
+#include <tuple>
+#include <utility>
+
+namespace at {
+
+TORCH_API std::vector<int64_t> infer_size(IntArrayRef a, IntArrayRef b);
+TORCH_API std::vector<SymInt> infer_size_symint(
+    SymIntArrayRef a,
+    SymIntArrayRef b);
+TORCH_API DimVector infer_size_dimvector(IntArrayRef a, IntArrayRef b);
+TORCH_API SymDimVector
+infer_size_symdimvector(SymIntArrayRef a, SymIntArrayRef b);
+
+// Named type instead of a pair/tuple so that we can be sure to
+// construct the vectors in place and get NRVO.
+template <typename Container>
+struct InferExpandGeometryResult {
+  Container sizes;
+  Container strides;
+  explicit InferExpandGeometryResult(size_t ndim)
+      : sizes(ndim), strides(ndim) {}
+  explicit InferExpandGeometryResult(IntArrayRef sizes_, size_t ndim)
+      : sizes(sizes_.begin(), sizes_.end()), strides(ndim) {}
+};
+
+TORCH_API std::tuple<std::vector<int64_t>, std::vector<int64_t>>
+inferExpandGeometry(
+    IntArrayRef tensor_sizes,
+    IntArrayRef tensor_strides,
+    IntArrayRef sizes);
+
+TORCH_API InferExpandGeometryResult<DimVector> inferExpandGeometry_dimvector(
+    IntArrayRef tensor_sizes,
+    IntArrayRef tensor_strides,
+    IntArrayRef sizes);
+
+TORCH_API std::vector<int64_t> infer_dense_strides(
+    IntArrayRef tensor_sizes,
+    IntArrayRef tensor_strides);
+
+// True if input shapes are expandable
+// NOTE: infer_size did a similar check, please keep them sync if change is
+// needed
+inline bool are_expandable(IntArrayRef shape1, IntArrayRef shape2) {
+  size_t ndim1 = shape1.size();
+  size_t ndim2 = shape2.size();
+  size_t ndim = ndim1 < ndim2 ? ndim1 : ndim2;
+
+  for (int64_t i = static_cast<int64_t>(ndim) - 1; i >= 0; --i) {
+    if (shape1[--ndim1] == shape2[--ndim2] || shape1[ndim1] == 1 ||
+        shape2[ndim2] == 1) {
+      continue;
+    }
+    return false;
+  }
+  return true;
+}
+
+// avoid copy-construction of Tensor by using a reference_wrapper.
+inline void check_defined(
+    std::initializer_list<std::reference_wrapper<const Tensor>> tensors,
+    const char* api_name) {
+  for (auto& t : tensors) {
+    if (!t.get().defined()) {
+      AT_ERROR(api_name, "(...) called with an undefined Tensor");
+    }
+  }
+}
+
+// NOTE [ ExpandUtils Borrowing ]
+//
+// Functions in ExpandUtils return `c10::MaybeOwned<Tensor>` because
+// expansion may not actually be needed, in which case we can improve
+// efficiency by returning
+// `c10::MaybeOwned<Tensor>::borrowed(to_expand)`. However, this means
+// that you need to be careful: the returned `c10::MaybeOwned<Tensor>`
+// must not outlive the original `Tensor` object that `to_expand`
+// referred to! The deleted rvalue reference overloads of these
+// functions help with this by preventing trivial use of a temporary
+// resulting from a function call, but it is still possible to make a
+// mistake.
+
+inline c10::MaybeOwned<Tensor> expand_inplace(
+    const Tensor& tensor,
+    const Tensor& to_expand) {
+  if (tensor.sym_sizes().equals(to_expand.sym_sizes())) {
+    return c10::MaybeOwned<Tensor>::borrowed(to_expand);
+  }
+  return c10::MaybeOwned<Tensor>::owned(
+      to_expand.expand_symint(tensor.sym_sizes()));
+}
+
+inline c10::MaybeOwned<Tensor> expand_inplace(
+    const Tensor& tensor,
+    Tensor&& to_expand) = delete;
+
+inline c10::MaybeOwned<Tensor> expand_inplace(
+    const Tensor& tensor,
+    const Tensor& to_expand,
+    const char* api_name) {
+  check_defined({tensor, to_expand}, api_name);
+  return expand_inplace(tensor, to_expand);
+}
+
+inline c10::MaybeOwned<Tensor> expand_inplace(
+    const Tensor& tensor,
+    Tensor&& to_expand,
+    const char* api_name) = delete;
+
+inline std::tuple<c10::MaybeOwned<Tensor>, c10::MaybeOwned<Tensor>>
+expand_inplace(
+    const Tensor& tensor,
+    const Tensor& to_expand1,
+    const Tensor& to_expand2) {
+  if (tensor.sizes().equals(to_expand1.sizes()) &&
+      tensor.sizes().equals((to_expand2.sizes()))) {
+    return std::make_tuple(
+        c10::MaybeOwned<Tensor>::borrowed(to_expand1),
+        c10::MaybeOwned<Tensor>::borrowed(to_expand2));
+  }
+
+  return std::make_tuple(
+      c10::MaybeOwned<Tensor>::owned(to_expand1.expand(tensor.sizes())),
+      c10::MaybeOwned<Tensor>::owned(to_expand2.expand(tensor.sizes())));
+}
+
+inline std::tuple<c10::MaybeOwned<Tensor>, c10::MaybeOwned<Tensor>>
+expand_inplace(
+    const Tensor& tensor,
+    Tensor&& to_expand1,
+    const Tensor& to_expand2) = delete;
+inline std::tuple<c10::MaybeOwned<Tensor>, c10::MaybeOwned<Tensor>>
+expand_inplace(
+    const Tensor& tensor,
+    const Tensor& to_expand1,
+    Tensor&& to_expand2) = delete;
+inline std::tuple<c10::MaybeOwned<Tensor>, c10::MaybeOwned<Tensor>>
+expand_inplace(const Tensor& tensor, Tensor&& to_expand1, Tensor&& to_expand2) =
+    delete;
+
+inline std::tuple<c10::MaybeOwned<Tensor>, c10::MaybeOwned<Tensor>>
+expand_inplace(
+    const Tensor& tensor,
+    const Tensor& to_expand1,
+    const Tensor& to_expand2,
+    const char* api_name) {
+  check_defined({tensor, to_expand1, to_expand2}, api_name);
+  return expand_inplace(tensor, to_expand1, to_expand2);
+}
+
+inline std::tuple<c10::MaybeOwned<Tensor>, c10::MaybeOwned<Tensor>>
+expand_inplace(
+    const Tensor& tensor,
+    Tensor&& to_expand1,
+    const Tensor& to_expand2,
+    const char* api_name) = delete;
+inline std::tuple<c10::MaybeOwned<Tensor>, c10::MaybeOwned<Tensor>>
+expand_inplace(
+    const Tensor& tensor,
+    const Tensor& to_expand1,
+    Tensor&& to_expand2,
+    const char* api_name) = delete;
+inline std::tuple<c10::MaybeOwned<Tensor>, c10::MaybeOwned<Tensor>>
+expand_inplace(
+    const Tensor& tensor,
+    Tensor&& to_expand1,
+    Tensor&& to_expand2,
+    const char* api_name) = delete;
+
+// See NOTE [ ExpandUtils Borrowing ] above for `MaybeOwned` explanation.
+inline std::tuple<c10::MaybeOwned<Tensor>, c10::MaybeOwned<Tensor>>
+expand_outplace(const Tensor& to_expand1, const Tensor& to_expand2) {
+  auto s1 = to_expand1.sym_sizes();
+  auto s2 = to_expand2.sym_sizes();
+  if (s1.equals(s2)) {
+    return std::make_tuple(
+        c10::MaybeOwned<Tensor>::borrowed(to_expand1),
+        c10::MaybeOwned<Tensor>::borrowed(to_expand2));
+  }
+
+  auto expanded_size = infer_size_symdimvector(s1, s2);
+  return std::make_tuple(
+      c10::MaybeOwned<Tensor>::owned(to_expand1.expand_symint(expanded_size)),
+      c10::MaybeOwned<Tensor>::owned(to_expand2.expand_symint(expanded_size)));
+}
+
+inline std::tuple<c10::MaybeOwned<Tensor>, c10::MaybeOwned<Tensor>>
+expand_outplace(Tensor&& to_expand1, const Tensor& to_expand2) = delete;
+inline std::tuple<c10::MaybeOwned<Tensor>, c10::MaybeOwned<Tensor>>
+expand_outplace(const Tensor& to_expand1, Tensor&& to_expand2) = delete;
+inline std::tuple<c10::MaybeOwned<Tensor>, c10::MaybeOwned<Tensor>>
+expand_outplace(Tensor&& to_expand1, Tensor&& to_expand2) = delete;
+
+inline std::tuple<c10::MaybeOwned<Tensor>, c10::MaybeOwned<Tensor>>
+expand_outplace(
+    const Tensor& to_expand1,
+    const Tensor& to_expand2,
+    const char* api_name) {
+  check_defined({to_expand1, to_expand2}, api_name);
+  return expand_outplace(to_expand1, to_expand2);
+}
+
+inline std::tuple<c10::MaybeOwned<Tensor>, c10::MaybeOwned<Tensor>>
+expand_outplace(
+    Tensor&& to_expand1,
+    const Tensor& to_expand2,
+    const char* api_name) = delete;
+inline std::tuple<c10::MaybeOwned<Tensor>, c10::MaybeOwned<Tensor>>
+expand_outplace(
+    const Tensor& to_expand1,
+    Tensor&& to_expand2,
+    const char* api_name) = delete;
+inline std::tuple<c10::MaybeOwned<Tensor>, c10::MaybeOwned<Tensor>>
+expand_outplace(
+    Tensor&& to_expand1,
+    Tensor&& to_expand2,
+    const char* api_name) = delete;
+
+inline std::tuple<
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>>
+expand_outplace(
+    const Tensor& to_expand1,
+    const Tensor& to_expand2,
+    const Tensor& to_expand3) {
+  if (to_expand1.sizes().equals(to_expand2.sizes()) &&
+      to_expand1.sizes().equals(to_expand3.sizes())) {
+    return std::make_tuple(
+        c10::MaybeOwned<Tensor>::borrowed(to_expand1),
+        c10::MaybeOwned<Tensor>::borrowed(to_expand2),
+        c10::MaybeOwned<Tensor>::borrowed(to_expand3));
+  }
+
+  auto expanded_size12 =
+      infer_size_dimvector(to_expand1.sizes(), to_expand2.sizes());
+  auto expanded_size =
+      infer_size_dimvector(expanded_size12, to_expand3.sizes());
+  return std::make_tuple(
+      c10::MaybeOwned<Tensor>::owned(to_expand1.expand(expanded_size)),
+      c10::MaybeOwned<Tensor>::owned(to_expand2.expand(expanded_size)),
+      c10::MaybeOwned<Tensor>::owned(to_expand3.expand(expanded_size)));
+}
+
+inline std::tuple<
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>>
+expand_outplace(
+    Tensor&& to_expand1,
+    const Tensor& to_expand2,
+    const Tensor& to_expand3) = delete;
+inline std::tuple<
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>>
+expand_outplace(
+    const Tensor& to_expand1,
+    Tensor&& to_expand2,
+    const Tensor& to_expand3) = delete;
+inline std::tuple<
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>>
+expand_outplace(
+    Tensor&& to_expand1,
+    Tensor&& to_expand2,
+    const Tensor& to_expand3) = delete;
+inline std::tuple<
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>>
+expand_outplace(
+    const Tensor& to_expand1,
+    const Tensor& to_expand2,
+    Tensor&& to_expand3) = delete;
+inline std::tuple<
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>>
+expand_outplace(
+    Tensor&& to_expand1,
+    const Tensor& to_expand2,
+    Tensor&& to_expand3) = delete;
+inline std::tuple<
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>>
+expand_outplace(
+    const Tensor& to_expand1,
+    Tensor&& to_expand2,
+    Tensor&& to_expand3) = delete;
+inline std::tuple<
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>>
+expand_outplace(Tensor&& to_expand1, Tensor&& to_expand2, Tensor&& to_expand3) =
+    delete;
+
+inline std::tuple<
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>>
+expand_outplace(
+    const Tensor& to_expand1,
+    const Tensor& to_expand2,
+    const Tensor& to_expand3,
+    const char* api_name) {
+  check_defined({to_expand1, to_expand2, to_expand3}, api_name);
+  return expand_outplace(to_expand1, to_expand2, to_expand3);
+}
+
+inline std::tuple<
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>>
+expand_outplace(
+    Tensor&& to_expand1,
+    const Tensor& to_expand2,
+    const Tensor& to_expand3,
+    const char* api_name) = delete;
+inline std::tuple<
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>>
+expand_outplace(
+    const Tensor& to_expand1,
+    Tensor&& to_expand2,
+    const Tensor& to_expand3,
+    const char* api_name) = delete;
+inline std::tuple<
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>>
+expand_outplace(
+    Tensor&& to_expand1,
+    Tensor&& to_expand2,
+    const Tensor& to_expand3,
+    const char* api_name) = delete;
+inline std::tuple<
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>>
+expand_outplace(
+    const Tensor& to_expand1,
+    const Tensor& to_expand2,
+    Tensor&& to_expand3,
+    const char* api_name) = delete;
+inline std::tuple<
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>>
+expand_outplace(
+    Tensor&& to_expand1,
+    const Tensor& to_expand2,
+    Tensor&& to_expand3,
+    const char* api_name) = delete;
+inline std::tuple<
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>>
+expand_outplace(
+    const Tensor& to_expand1,
+    Tensor&& to_expand2,
+    Tensor&& to_expand3,
+    const char* api_name) = delete;
+inline std::tuple<
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>,
+    c10::MaybeOwned<Tensor>>
+expand_outplace(
+    Tensor&& to_expand1,
+    Tensor&& to_expand2,
+    Tensor&& to_expand3,
+    const char* api_name) = delete;
+
+inline c10::MaybeOwned<Tensor> expand_size(
+    const Tensor& to_expand,
+    IntArrayRef sizes) {
+  if (to_expand.sizes().equals(sizes)) {
+    return c10::MaybeOwned<Tensor>::borrowed(to_expand);
+  }
+
+  return c10::MaybeOwned<Tensor>::owned(to_expand.expand(sizes));
+}
+
+inline c10::MaybeOwned<Tensor> expand_size(
+    Tensor&& to_expand,
+    IntArrayRef sizes) = delete;
+
+inline c10::MaybeOwned<Tensor> expand_size(
+    const Tensor& to_expand,
+    IntArrayRef sizes,
+    const char* api_name) {
+  check_defined({to_expand}, api_name);
+  return expand_size(to_expand, sizes);
+}
+
+inline c10::MaybeOwned<Tensor> expand_size(
+    Tensor&& to_expand,
+    IntArrayRef sizes,
+    const char* api_name) = delete;
+
+inline std::vector<Tensor> expand_outplace(TensorList to_expand) {
+  // expands a list of Tensors; ignores undefined (null) tensors
+  bool first = true;
+  DimVector sizes;
+  for (const auto i : c10::irange(to_expand.size())) {
+    if (!to_expand[i].defined()) {
+      continue;
+    } else if (first) {
+      sizes = to_expand[i].sizes();
+      first = false;
+    } else {
+      sizes = infer_size_dimvector(sizes, to_expand[i].sizes());
+    }
+  }
+
+  std::vector<Tensor> result(to_expand.size());
+  for (const auto i : c10::irange(to_expand.size())) {
+    if (!to_expand[i].defined()) {
+      continue;
+    } else if (to_expand[i].sizes().equals(sizes)) {
+      result[i] = to_expand[i];
+    } else {
+      result[i] = to_expand[i].expand(sizes);
+    }
+  }
+  return result;
+}
+
+template <typename T>
+inline Tensor _sum_to(
+    Tensor tensor,
+    const c10::ArrayRef<T> shape,
+    bool always_return_non_view = false) {
+  if (shape.size() == 0) {
+    return tensor.sum();
+  }
+
+  auto sizes = at::symint::sizes<T>(tensor);
+  c10::SmallVector<int64_t, 8> reduce_dims;
+  const int64_t leading_dims = sizes.size() - shape.size();
+  for (const auto i : c10::irange(leading_dims)) {
+    reduce_dims.push_back(i);
+  }
+  for (int64_t i = leading_dims; i < static_cast<int64_t>(sizes.size()); ++i) {
+    if (shape[i - leading_dims] == 1 && sizes[i] != 1) {
+      reduce_dims.push_back(i);
+    }
+  }
+
+  if (!reduce_dims.empty()) {
+    tensor = tensor.sum(reduce_dims, /*keepdim=*/true);
+  }
+
+  if (always_return_non_view) {
+    // This is only actually used by the functionalization pass.
+    // We want to be able to guarantee that this function doesn't return a view
+    // of the input.
+    return leading_dims > 0 ? at::symint::view_copy<T>(tensor, shape)
+                            : tensor.clone();
+  } else {
+    return leading_dims > 0 ? at::symint::view<T>(tensor, shape) : tensor;
+  }
+}
+
+inline Tensor sum_to(
+    Tensor tensor,
+    const c10::SymIntArrayRef shape,
+    bool always_return_non_view = false) {
+  return _sum_to(std::move(tensor), shape, always_return_non_view);
+}
+
+// Sums `tensor` repeatedly to produce a tensor of shape `shape`.
+// Precondition: is_expandable_to(shape, tensor.sizes()) must be true
+inline Tensor sum_to(
+    Tensor tensor,
+    const IntArrayRef shape,
+    bool always_return_non_view = false) {
+  return _sum_to(std::move(tensor), shape, always_return_non_view);
+}
+
+static inline bool is_expandable_to(
+    SymIntArrayRef shape,
+    c10::SymIntArrayRef desired) {
+  size_t ndim = shape.size();
+  size_t target_dim = desired.size();
+  if (ndim > target_dim) {
+    return false;
+  }
+  for (const auto i : c10::irange(ndim)) {
+    const auto& size = shape[ndim - i - 1];
+    const auto& target = desired[target_dim - i - 1];
+    if (size != target && size != 1) {
+      return false;
+    }
+  }
+  return true;
+}
+
+static inline bool is_expandable_to(IntArrayRef shape, IntArrayRef desired) {
+  auto sym_shape = c10::SymIntArrayRef(
+      reinterpret_cast<const c10::SymInt*>(shape.data()), shape.size());
+  auto sym_desired = c10::SymIntArrayRef(
+      reinterpret_cast<const c10::SymInt*>(desired.data()), desired.size());
+  return is_expandable_to(sym_shape, sym_desired);
+}
+
+} // namespace at

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

@@ -0,0 +1,126 @@
+#pragma once
+
+#include <ATen/Tensor.h>
+
+namespace at::functionalization {
+
+// See Note [Functionalization Pass In Core]
+
+// ViewMeta is a class used by the functionalization pass to navigate between
+// a base tensor and a view tensor.
+// For example, if I call `b = a.view1(...)`
+// the functionalization pass will generate and store a ViewMeta on b that looks
+// like:
+//
+// ViewMeta(
+//   [<captures>](const Tensor& base, int64_t mutated_view_idx) {
+//     return base.view1(...);
+//   },
+//   [<captures>](const at::Tensor& base, const at::Tensor& mutated_view,
+//   int64_t mutated_view_idx) -> at::Tensor {
+//     return at::functionalization::impl::view1_inverse(base, mutated_view,
+//     ...);
+//   }
+//
+// The forward_fn lambda describes how to replay view1 on a tensor.
+//
+// The reverse_fn lambda describes how, given a tensor that is already a view,
+// how to get the corresponding base tensor. See Note [Functionalization Pass:
+// View Inverses] for details.
+struct ViewMeta {
+  ViewMeta(
+      std::function<Tensor(const Tensor&, int64_t)> forward,
+      std::function<Tensor(const Tensor&, const Tensor&, int64_t)> reverse,
+      bool is_multi_output = false,
+      int64_t out_idx = 0)
+      : forward_fn(std::move(forward)),
+        reverse_fn(std::move(reverse)),
+        out_index(out_idx),
+        is_multi_output(is_multi_output) {}
+
+  std::function<Tensor(const Tensor&, int64_t)> forward_fn;
+  std::function<Tensor(const Tensor&, const Tensor&, int64_t)> reverse_fn;
+  // See Note [out_idx in ViewMeta]
+  int64_t out_index;
+
+  // Tells us if this is a multi-output view
+  bool is_multi_output;
+
+  // Returns a copy of the current ViewMeta, if out_idx matches the current
+  // out_index. Otherwise, returns a new ViewMeta with the same forward/reverse
+  // functions, but a new out index.
+  ViewMeta to_out_idx(int64_t out_idx);
+};
+
+// FunctionalStorageImpl is a subclass of StorageImpl used by the
+// functionalization pass. It has no underlying data (similar to meta storage).
+// It also knows how to reflect mutations to tensors in the absence of a valid
+// data pointer.
+//
+// A storage represents the state shared by (potentially multiple) views of the
+// same tensor. For example, in the following code:
+//
+// b = a.view1(...)
+// c = b.view2(...)
+// b.add_(1)
+// --> storage.add_update(b, {view1_meta})
+//
+// The call to add_(1) will result in a call to alias.add_update(b,
+// {view1_meta}), queueing up the mutation from b onto the alias. Later, suppose
+// c is used in an expression (e.g. you try to print c, or pass it to an
+// operator). Doing so will involve "syncing" c. First we apply any pending
+// updates to the alias, and then we regenerate c by replaying its views off of
+// the updated alias. E.g:
+//
+// print(str(c))
+// --> c.sync_()
+//     --> alias.apply_updates() // after this, the alias will be updated to
+//     reflect the mutation to b
+struct TORCH_API FunctionalStorageImpl : public c10::StorageImpl {
+ public:
+  struct Update {
+    // NOLINTNEXTLINE(cppcoreguidelines-avoid-const-or-ref-data-members)
+    const at::Tensor new_val;
+    // NOLINTNEXTLINE(cppcoreguidelines-avoid-const-or-ref-data-members)
+    const std::vector<ViewMeta> view_metas;
+  };
+
+  explicit FunctionalStorageImpl(const Tensor& value);
+
+  void add_update(
+      const Tensor& updated_val,
+      const std::vector<ViewMeta>& view_metas);
+  bool apply_updates();
+  const Tensor& base() {
+    return base_;
+  }
+  size_t generation() const {
+    return generation_;
+  }
+  void freeze() {
+    frozen_ = true;
+  }
+
+  ~FunctionalStorageImpl() override = default;
+
+ private:
+  // NB: base_ should always point to a tensor BELOW the current
+  // functionalization layer. This is mainly to avoid reference cycles. e.g.
+  // given `b = a.view(...)` Both a.storage_ and b.storage_ are a
+  // FunctionStorageImpl containing an Walualias, with contains a Tensor
+  // `base_`. In this case (where a and b are FunctionalTensorWrapper's), base_
+  // should point not to a, but to a's unwrapped value, a.value_` See Note
+  // [Functionalization: Walualias Removal] for a diagram that shows this
+  // visually.
+  at::Tensor base_;
+  std::vector<Update> updates_;
+  // generation_ gets incremented every time a mutation is queued onto the
+  // alias. It is used to determine if a given tensor is "up to date", or if it
+  // needs to be regenerated from the alias.
+  size_t generation_ = 0;
+  // If frozen, no more mutations are allowed on this storage.  Once frozen, a
+  // storage cannot be unfrozen.
+  bool frozen_ = false;
+};
+
+} // namespace at::functionalization

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

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

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

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

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/LinalgBackend.h

@@ -0,0 +1,31 @@
+#pragma once
+
+#include <c10/util/Exception.h>
+
+#include <ostream>
+#include <string>
+
+namespace at {
+
+enum class LinalgBackend : int8_t { Default, Cusolver, Magma };
+
+inline std::string LinalgBackendToString(at::LinalgBackend backend) {
+  switch (backend) {
+    case LinalgBackend::Default:
+      return "at::LinalgBackend::Default";
+    case LinalgBackend::Cusolver:
+      return "at::LinalgBackend::Cusolver";
+    case LinalgBackend::Magma:
+      return "at::LinalgBackend::Magma";
+    default:
+      TORCH_CHECK(false, "Unknown linalg backend");
+  }
+}
+
+inline std::ostream& operator<<(
+    std::ostream& stream,
+    at::LinalgBackend backend) {
+  return stream << LinalgBackendToString(backend);
+}
+
+} // namespace at

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

@@ -0,0 +1,109 @@
+#pragma once
+#include <ATen/Utils.h>
+#include <c10/util/ArrayRef.h>
+
+#include <vector>
+
+namespace at {
+/// MatrixRef - Like an ArrayRef, but with an extra recorded strides so that
+/// we can easily view it as a multidimensional array.
+///
+/// Like ArrayRef, this class does not own the underlying data, it is expected
+/// to be used in situations where the data resides in some other buffer.
+///
+/// This is intended to be trivially copyable, so it should be passed by
+/// value.
+///
+/// For now, 2D only (so the copies are actually cheap, without having
+/// to write a SmallVector class) and contiguous only (so we can
+/// return non-strided ArrayRef on index).
+///
+/// P.S. dimension 0 indexes rows, dimension 1 indexes columns
+template <typename T>
+class MatrixRef {
+ public:
+  typedef size_t size_type;
+
+ private:
+  /// Underlying ArrayRef
+  ArrayRef<T> arr;
+
+  /// Stride of dim 0 (outer dimension)
+  size_type stride0;
+
+  // Stride of dim 1 is assumed to be 1
+
+ public:
+  /// Construct an empty Matrixref.
+  /*implicit*/ MatrixRef() : arr(nullptr), stride0(0) {}
+
+  /// Construct an MatrixRef from an ArrayRef and outer stride.
+  /*implicit*/ MatrixRef(ArrayRef<T> arr, size_type stride0)
+      : arr(arr), stride0(stride0) {
+    TORCH_CHECK(
+        arr.size() % stride0 == 0,
+        "MatrixRef: ArrayRef size ",
+        arr.size(),
+        " not divisible by stride ",
+        stride0)
+  }
+
+  /// @}
+  /// @name Simple Operations
+  /// @{
+
+  /// empty - Check if the matrix is empty.
+  bool empty() const {
+    return arr.empty();
+  }
+
+  const T* data() const {
+    return arr.data();
+  }
+
+  /// size - Get size a dimension
+  size_t size(size_t dim) const {
+    if (dim == 0) {
+      return arr.size() / stride0;
+    } else if (dim == 1) {
+      return stride0;
+    } else {
+      TORCH_CHECK(
+          0, "MatrixRef: out of bounds dimension ", dim, "; expected 0 or 1");
+    }
+  }
+
+  size_t numel() const {
+    return arr.size();
+  }
+
+  /// equals - Check for element-wise equality.
+  bool equals(MatrixRef RHS) const {
+    return stride0 == RHS.stride0 && arr.equals(RHS.arr);
+  }
+
+  /// @}
+  /// @name Operator Overloads
+  /// @{
+  ArrayRef<T> operator[](size_t Index) const {
+    return arr.slice(Index * stride0, stride0);
+  }
+
+  /// Disallow accidental assignment from a temporary.
+  ///
+  /// The declaration here is extra complicated so that "arrayRef = {}"
+  /// continues to select the move assignment operator.
+  template <typename U>
+  std::enable_if_t<std::is_same_v<U, T>, MatrixRef<T>>& operator=(
+      U&& Temporary) = delete;
+
+  /// Disallow accidental assignment from a temporary.
+  ///
+  /// The declaration here is extra complicated so that "arrayRef = {}"
+  /// continues to select the move assignment operator.
+  template <typename U>
+  std::enable_if_t<std::is_same_v<U, T>, MatrixRef<T>>& operator=(
+      std::initializer_list<U>) = delete;
+};
+
+} // end namespace at

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

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

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

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

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

@@ -0,0 +1,215 @@
+#pragma once
+#include <ATen/NamedTensor.h>
+#include <ATen/TensorNames.h>
+#include <ATen/WrapDimUtilsMulti.h>
+
+#include <ATen/core/DimVector.h>
+#include <ATen/core/Tensor.h>
+#include <functional>
+
+namespace at {
+
+using NameVector = SmallVector<Dimname, kDimVectorStaticSize>;
+
+inline bool has_names(const ITensorListRef& tensors) {
+  return std::any_of(tensors.begin(), tensors.end(), [](const Tensor& t) {
+    return t.has_names();
+  });
+}
+
+// Converts dim to an positional index. Errors if `dim` cannot be used to
+// refer to any dimension of tensor.
+TORCH_API int64_t dimname_to_position(const Tensor& tensor, Dimname dim);
+TORCH_API std::vector<int64_t> dimnames_to_positions(
+    const Tensor& tensor,
+    DimnameList dims);
+
+// Unifies two DimnameList to produce a third. This is useful for implementing
+// the named inference rule for binary broadcasting operations like add.
+//
+// There are three main constraints:
+// 1) Check matching: Names must match positionally from the right.
+// 2) Check misaligned: If a name `n` is in `names`, then it must appear at
+//    the same index from the right in other.
+// 3) The output names are obtained by unifying the names individually from the
+// right.
+TORCH_API std::vector<Dimname> unify_from_right(
+    DimnameList names,
+    DimnameList other,
+    const char* action = "broadcast");
+
+[[noreturn]] inline void reportNYIDimnameOverload(const char* op_name) {
+  TORCH_CHECK(
+      false,
+      op_name,
+      ": You passed a dimname (string) to this op in place of a dimension "
+      "index but it does not yet support this behavior. Please pass a dimension "
+      "index to work around this.");
+}
+
+// [NOTE] Writing name inference rules
+//
+// Operators that support named tensors are either composed of operations that
+// support named tensors or implement some name inference rule. An op that
+// implements its own name inference rule generally looks like the following:
+//
+// Tensor op(...) {
+//   perform_shape_checks(...);
+//   # (1)
+//   auto maybe_outnames = compute_outnames(...);
+//   auto result = [&]() {
+//     NoNamesGuard guard;
+//     return op_impl(...);
+//   }();
+//   # (2)
+//   propagate_names_if_nonempty(result, maybe_outnames);
+//
+// Each op has (1) a compute outnames step and (2) a propagate names step.
+//
+// compute_outnames is responsible for checking that input names match and
+// determining what the output names should be. It returns either:
+// - {} (if the inputs tensors are all unnamed)
+// - non-empty outnames.
+//
+// propagate_names_if_nonempty propagates the outnames if they exist to the
+// result tensors.
+//
+// The {} case is an optimization; if the user does not use named tensors they
+// pay no perf cost for it.
+
+namespace namedinference {
+
+const Tensor& propagate_names_if_present_and_nonempty(
+    const Tensor& result,
+    c10::optional<DimnameList> maybe_names,
+    bool validate_names = false);
+// Propagates `names` to `result` if `names` is not empty.
+// `names` can be empty; see [NOTE] Writing name inference rules
+// If `names` is not empty, `names.size()` should equal `result.dim()`.
+// When in doubt, use this overload instead of the others.
+TORCH_API const Tensor& propagate_names_if_nonempty(
+    const Tensor& result,
+    DimnameList maybe_names,
+    bool validate_names = false);
+
+// Propagates `names` to `result`. Only use this if we are certain that there
+// are names to propagate (that names is not empty).
+TORCH_API const Tensor& propagate_names(
+    const Tensor& result,
+    DimnameList names,
+    bool validate_names = false);
+
+// Propagates all names from src to result.
+TORCH_API void propagate_names(const Tensor& result, const Tensor& src);
+
+// Propagates all names except for those at the excluded_idxs.
+TORCH_API void propagate_names_except(
+    const Tensor& result,
+    const Tensor& src,
+    IntArrayRef excluded_idxs);
+
+// Used for reduction ops that have a `keepdim` arg.
+TORCH_API void propagate_names_for_reduction(
+    const Tensor& result,
+    const Tensor& src,
+    IntArrayRef excluded_idxs,
+    bool keepdim);
+
+TORCH_API void propagate_names_for_expand(
+    const Tensor& result,
+    const Tensor& self);
+
+TORCH_API std::vector<Dimname> compute_cat_outnames(
+    const MaterializedITensorListRef& tensors);
+
+TORCH_API std::vector<Dimname> compute_broadcast_outnames(
+    const Tensor& self,
+    const Tensor& other);
+
+TORCH_API std::vector<Dimname> broadcast_to_outnames(
+    const Tensor& tensor,
+    const Tensor& reference_tensor,
+    const char* op_name);
+
+TORCH_API std::vector<Dimname> compute_matmul_outnames(
+    const Tensor& self,
+    const Tensor& other);
+
+TORCH_API std::vector<Dimname> compute_cdist_outnames(
+    const Tensor& self,
+    const Tensor& other);
+
+TORCH_API std::vector<Dimname> compute_bmm_outnames(
+    const Tensor& result,
+    const Tensor& self,
+    const Tensor& other);
+
+TORCH_API std::vector<Dimname> compute_squeeze_outnames(const Tensor& tensor);
+TORCH_API std::vector<Dimname> compute_squeeze_outnames(
+    const Tensor& tensor,
+    std::bitset<dim_bitset_size> dims);
+
+std::vector<Dimname> compute_diagonal_outnames(
+    const Tensor& tensor,
+    int64_t dim1,
+    int64_t dim2);
+
+// TensorImpl* overloads for Legacy TH/THC code. Use these sparingly.
+
+TORCH_API TensorImpl* propagate_names_if_nonempty(
+    TensorImpl* result,
+    DimnameList maybe_names,
+    bool validate_names = false);
+
+TORCH_API TensorImpl* propagate_names(
+    TensorImpl* result,
+    DimnameList names,
+    bool validate_names = false);
+
+TORCH_API void propagate_names(TensorImpl* result, /*const */ TensorImpl* src);
+
+TORCH_API inline void propagate_names(
+    const TensorBase& result,
+    DimnameList names,
+    bool validate_names = false) {
+  propagate_names(result.unsafeGetTensorImpl(), names, validate_names);
+}
+
+TORCH_API inline void propagate_names_if_nonempty(
+    const TensorBase& result,
+    DimnameList names,
+    bool validate_names = false) {
+  propagate_names_if_nonempty(
+      result.unsafeGetTensorImpl(), names, validate_names);
+}
+
+TORCH_API inline void propagate_names(
+    const TensorBase& result,
+    const TensorBase& src) {
+  propagate_names(result.unsafeGetTensorImpl(), src.unsafeGetTensorImpl());
+}
+
+// result = m1 @ m2 + bias
+TORCH_API std::vector<Dimname> propagate_names_for_addmm(
+    const Tensor& m1,
+    const Tensor& m2,
+    const Tensor& bias);
+
+TORCH_API std::vector<Dimname> propagate_names_for_addmv(
+    const Tensor& mat,
+    const Tensor& vec,
+    const Tensor& bias);
+
+TORCH_API void check_names_for_dot(TensorImpl* vec1, TensorImpl* vec2);
+
+TORCH_API std::vector<Dimname> compute_baddbmm_outnames(
+    const Tensor& result,
+    const Tensor& self,
+    const Tensor& other,
+    const Tensor& bias);
+
+TORCH_API bool are_names_equal(TensorImpl* self, TensorImpl* other);
+
+} // namespace namedinference
+
+} // namespace at

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

@@ -0,0 +1,1317 @@
+#pragma once
+
+// @generated by torchgen/gen.py from NativeFunctions.h
+
+#ifdef TORCH_ASSERT_NO_OPERATORS
+#error This change adds a dependency on native_functions.yaml,            \
+  meaning the file will need to be re-compiled every time an operator     \
+  is changed or added. Consider if your change would be better placed in  \
+  another file, or if a more specific header might achieve the same goal. \
+  See NOTE: [Tensor vs. TensorBase]
+#endif
+
+#if defined(AT_PER_OPERATOR_HEADERS) && defined(TORCH_ASSERT_ONLY_METHOD_OPERATORS)
+#error This change adds a dependency on all pytorch operators, meaning the      \
+  file will need to be re-compiled every time an operator is changed or added.  \
+  Consider including a specific operator from <ATen/ops/{my_operator}_native.h> \
+  and see NOTE [TORCH_ASSERT_ONLY_METHOD_OPERATORS].
+#endif
+
+#include <c10/core/Scalar.h>
+#include <c10/core/Storage.h>
+#include <c10/core/TensorOptions.h>
+#include <c10/util/Deprecated.h>
+#include <c10/util/Optional.h>
+#include <c10/core/QScheme.h>
+#include <ATen/core/Reduction.h>
+#include <ATen/core/Tensor.h>
+#include <tuple>
+#include <vector>
+
+#include <ATen/ops/_adaptive_avg_pool2d_native.h>
+#include <ATen/ops/_adaptive_avg_pool2d_backward_native.h>
+#include <ATen/ops/_adaptive_avg_pool3d_native.h>
+#include <ATen/ops/_adaptive_avg_pool3d_backward_native.h>
+#include <ATen/ops/_add_batch_dim_native.h>
+#include <ATen/ops/_add_relu_native.h>
+#include <ATen/ops/_addmm_activation_native.h>
+#include <ATen/ops/_aminmax_native.h>
+#include <ATen/ops/_amp_foreach_non_finite_check_and_unscale_native.h>
+#include <ATen/ops/_amp_update_scale_native.h>
+#include <ATen/ops/_assert_async_native.h>
+#include <ATen/ops/_assert_scalar_native.h>
+#include <ATen/ops/_assert_tensor_metadata_native.h>
+#include <ATen/ops/_autocast_to_full_precision_native.h>
+#include <ATen/ops/_autocast_to_reduced_precision_native.h>
+#include <ATen/ops/_backward_native.h>
+#include <ATen/ops/_batch_norm_impl_index_native.h>
+#include <ATen/ops/_batch_norm_impl_index_backward_native.h>
+#include <ATen/ops/_cast_Byte_native.h>
+#include <ATen/ops/_cast_Char_native.h>
+#include <ATen/ops/_cast_Double_native.h>
+#include <ATen/ops/_cast_Float_native.h>
+#include <ATen/ops/_cast_Half_native.h>
+#include <ATen/ops/_cast_Int_native.h>
+#include <ATen/ops/_cast_Long_native.h>
+#include <ATen/ops/_cast_Short_native.h>
+#include <ATen/ops/_cdist_backward_native.h>
+#include <ATen/ops/_cdist_forward_native.h>
+#include <ATen/ops/_cholesky_solve_helper_native.h>
+#include <ATen/ops/_choose_qparams_per_tensor_native.h>
+#include <ATen/ops/_chunk_cat_native.h>
+#include <ATen/ops/_coalesce_native.h>
+#include <ATen/ops/_coalesced_native.h>
+#include <ATen/ops/_compute_linear_combination_native.h>
+#include <ATen/ops/_conj_native.h>
+#include <ATen/ops/_conj_copy_native.h>
+#include <ATen/ops/_conj_physical_native.h>
+#include <ATen/ops/_conv_depthwise2d_native.h>
+#include <ATen/ops/_convert_indices_from_coo_to_csr_native.h>
+#include <ATen/ops/_convert_indices_from_csr_to_coo_native.h>
+#include <ATen/ops/_convert_weight_to_int4pack_native.h>
+#include <ATen/ops/_convolution_native.h>
+#include <ATen/ops/_convolution_double_backward_native.h>
+#include <ATen/ops/_convolution_mode_native.h>
+#include <ATen/ops/_copy_from_native.h>
+#include <ATen/ops/_copy_from_and_resize_native.h>
+#include <ATen/ops/_cslt_compress_native.h>
+#include <ATen/ops/_cslt_sparse_mm_native.h>
+#include <ATen/ops/_cslt_sparse_mm_search_native.h>
+#include <ATen/ops/_ctc_loss_native.h>
+#include <ATen/ops/_ctc_loss_backward_native.h>
+#include <ATen/ops/_cudnn_ctc_loss_native.h>
+#include <ATen/ops/_cudnn_init_dropout_state_native.h>
+#include <ATen/ops/_cudnn_rnn_native.h>
+#include <ATen/ops/_cudnn_rnn_backward_native.h>
+#include <ATen/ops/_cudnn_rnn_flatten_weight_native.h>
+#include <ATen/ops/_cufft_clear_plan_cache_native.h>
+#include <ATen/ops/_cufft_get_plan_cache_max_size_native.h>
+#include <ATen/ops/_cufft_get_plan_cache_size_native.h>
+#include <ATen/ops/_cufft_set_plan_cache_max_size_native.h>
+#include <ATen/ops/_cummax_helper_native.h>
+#include <ATen/ops/_cummin_helper_native.h>
+#include <ATen/ops/_debug_has_internal_overlap_native.h>
+#include <ATen/ops/_dimI_native.h>
+#include <ATen/ops/_dimV_native.h>
+#include <ATen/ops/_dim_arange_native.h>
+#include <ATen/ops/_dirichlet_grad_native.h>
+#include <ATen/ops/_efficient_attention_backward_native.h>
+#include <ATen/ops/_efficient_attention_forward_native.h>
+#include <ATen/ops/_efficientzerotensor_native.h>
+#include <ATen/ops/_embedding_bag_native.h>
+#include <ATen/ops/_embedding_bag_backward_native.h>
+#include <ATen/ops/_embedding_bag_dense_backward_native.h>
+#include <ATen/ops/_embedding_bag_forward_only_native.h>
+#include <ATen/ops/_embedding_bag_per_sample_weights_backward_native.h>
+#include <ATen/ops/_embedding_bag_sparse_backward_native.h>
+#include <ATen/ops/_empty_affine_quantized_native.h>
+#include <ATen/ops/_empty_per_channel_affine_quantized_native.h>
+#include <ATen/ops/_euclidean_dist_native.h>
+#include <ATen/ops/_fake_quantize_learnable_per_channel_affine_native.h>
+#include <ATen/ops/_fake_quantize_learnable_per_channel_affine_backward_native.h>
+#include <ATen/ops/_fake_quantize_learnable_per_tensor_affine_native.h>
+#include <ATen/ops/_fake_quantize_learnable_per_tensor_affine_backward_native.h>
+#include <ATen/ops/_fake_quantize_per_tensor_affine_cachemask_tensor_qparams_native.h>
+#include <ATen/ops/_fft_c2c_native.h>
+#include <ATen/ops/_fft_c2r_native.h>
+#include <ATen/ops/_fft_r2c_native.h>
+#include <ATen/ops/_fill_mem_eff_dropout_mask_native.h>
+#include <ATen/ops/_flash_attention_backward_native.h>
+#include <ATen/ops/_flash_attention_forward_native.h>
+#include <ATen/ops/_foobar_native.h>
+#include <ATen/ops/_foreach_abs_native.h>
+#include <ATen/ops/_foreach_acos_native.h>
+#include <ATen/ops/_foreach_add_native.h>
+#include <ATen/ops/_foreach_addcdiv_native.h>
+#include <ATen/ops/_foreach_addcmul_native.h>
+#include <ATen/ops/_foreach_asin_native.h>
+#include <ATen/ops/_foreach_atan_native.h>
+#include <ATen/ops/_foreach_ceil_native.h>
+#include <ATen/ops/_foreach_clamp_max_native.h>
+#include <ATen/ops/_foreach_clamp_min_native.h>
+#include <ATen/ops/_foreach_copy_native.h>
+#include <ATen/ops/_foreach_cos_native.h>
+#include <ATen/ops/_foreach_cosh_native.h>
+#include <ATen/ops/_foreach_div_native.h>
+#include <ATen/ops/_foreach_erf_native.h>
+#include <ATen/ops/_foreach_erfc_native.h>
+#include <ATen/ops/_foreach_exp_native.h>
+#include <ATen/ops/_foreach_expm1_native.h>
+#include <ATen/ops/_foreach_floor_native.h>
+#include <ATen/ops/_foreach_frac_native.h>
+#include <ATen/ops/_foreach_lerp_native.h>
+#include <ATen/ops/_foreach_lgamma_native.h>
+#include <ATen/ops/_foreach_log_native.h>
+#include <ATen/ops/_foreach_log10_native.h>
+#include <ATen/ops/_foreach_log1p_native.h>
+#include <ATen/ops/_foreach_log2_native.h>
+#include <ATen/ops/_foreach_maximum_native.h>
+#include <ATen/ops/_foreach_minimum_native.h>
+#include <ATen/ops/_foreach_mul_native.h>
+#include <ATen/ops/_foreach_neg_native.h>
+#include <ATen/ops/_foreach_norm_native.h>
+#include <ATen/ops/_foreach_pow_native.h>
+#include <ATen/ops/_foreach_reciprocal_native.h>
+#include <ATen/ops/_foreach_round_native.h>
+#include <ATen/ops/_foreach_sigmoid_native.h>
+#include <ATen/ops/_foreach_sign_native.h>
+#include <ATen/ops/_foreach_sin_native.h>
+#include <ATen/ops/_foreach_sinh_native.h>
+#include <ATen/ops/_foreach_sqrt_native.h>
+#include <ATen/ops/_foreach_sub_native.h>
+#include <ATen/ops/_foreach_tan_native.h>
+#include <ATen/ops/_foreach_tanh_native.h>
+#include <ATen/ops/_foreach_trunc_native.h>
+#include <ATen/ops/_foreach_zero_native.h>
+#include <ATen/ops/_functional_assert_async_native.h>
+#include <ATen/ops/_functional_assert_scalar_native.h>
+#include <ATen/ops/_functional_sym_constrain_range_native.h>
+#include <ATen/ops/_functional_sym_constrain_range_for_size_native.h>
+#include <ATen/ops/_fused_adam_native.h>
+#include <ATen/ops/_fused_adamw_native.h>
+#include <ATen/ops/_fused_dropout_native.h>
+#include <ATen/ops/_fused_moving_avg_obs_fq_helper_native.h>
+#include <ATen/ops/_fused_sdp_choice_native.h>
+#include <ATen/ops/_fused_sgd_native.h>
+#include <ATen/ops/_fw_primal_native.h>
+#include <ATen/ops/_fw_primal_copy_native.h>
+#include <ATen/ops/_gather_sparse_backward_native.h>
+#include <ATen/ops/_grid_sampler_2d_cpu_fallback_native.h>
+#include <ATen/ops/_grid_sampler_2d_cpu_fallback_backward_native.h>
+#include <ATen/ops/_has_compatible_shallow_copy_type_native.h>
+#include <ATen/ops/_has_same_storage_numel_native.h>
+#include <ATen/ops/_histogramdd_bin_edges_native.h>
+#include <ATen/ops/_histogramdd_from_bin_cts_native.h>
+#include <ATen/ops/_histogramdd_from_bin_tensors_native.h>
+#include <ATen/ops/_index_put_impl_native.h>
+#include <ATen/ops/_indices_native.h>
+#include <ATen/ops/_indices_copy_native.h>
+#include <ATen/ops/_int_mm_native.h>
+#include <ATen/ops/_is_all_true_native.h>
+#include <ATen/ops/_is_any_true_native.h>
+#include <ATen/ops/_is_zerotensor_native.h>
+#include <ATen/ops/_lazy_clone_native.h>
+#include <ATen/ops/_linalg_check_errors_native.h>
+#include <ATen/ops/_linalg_det_native.h>
+#include <ATen/ops/_linalg_eigh_native.h>
+#include <ATen/ops/_linalg_eigvals_native.h>
+#include <ATen/ops/_linalg_slogdet_native.h>
+#include <ATen/ops/_linalg_solve_ex_native.h>
+#include <ATen/ops/_linalg_svd_native.h>
+#include <ATen/ops/_local_scalar_dense_native.h>
+#include <ATen/ops/_log_softmax_native.h>
+#include <ATen/ops/_log_softmax_backward_data_native.h>
+#include <ATen/ops/_logcumsumexp_native.h>
+#include <ATen/ops/_lstm_mps_native.h>
+#include <ATen/ops/_lu_with_info_native.h>
+#include <ATen/ops/_make_dep_token_native.h>
+#include <ATen/ops/_make_dual_native.h>
+#include <ATen/ops/_make_dual_copy_native.h>
+#include <ATen/ops/_make_per_channel_quantized_tensor_native.h>
+#include <ATen/ops/_make_per_tensor_quantized_tensor_native.h>
+#include <ATen/ops/_masked_scale_native.h>
+#include <ATen/ops/_masked_softmax_native.h>
+#include <ATen/ops/_masked_softmax_backward_native.h>
+#include <ATen/ops/_mixed_dtypes_linear_native.h>
+#include <ATen/ops/_mkldnn_reshape_native.h>
+#include <ATen/ops/_mkldnn_transpose_native.h>
+#include <ATen/ops/_mps_convolution_native.h>
+#include <ATen/ops/_mps_convolution_transpose_native.h>
+#include <ATen/ops/_native_batch_norm_legit_native.h>
+#include <ATen/ops/_native_batch_norm_legit_no_training_native.h>
+#include <ATen/ops/_native_multi_head_attention_native.h>
+#include <ATen/ops/_neg_view_native.h>
+#include <ATen/ops/_neg_view_copy_native.h>
+#include <ATen/ops/_nested_from_padded_native.h>
+#include <ATen/ops/_nested_from_padded_and_nested_example_native.h>
+#include <ATen/ops/_nested_get_jagged_dummy_native.h>
+#include <ATen/ops/_nested_get_lengths_native.h>
+#include <ATen/ops/_nested_get_offsets_native.h>
+#include <ATen/ops/_nested_get_ragged_idx_native.h>
+#include <ATen/ops/_nested_get_values_native.h>
+#include <ATen/ops/_nested_get_values_copy_native.h>
+#include <ATen/ops/_nested_select_backward_native.h>
+#include <ATen/ops/_nested_sum_backward_native.h>
+#include <ATen/ops/_nested_tensor_from_mask_native.h>
+#include <ATen/ops/_nested_tensor_from_mask_left_aligned_native.h>
+#include <ATen/ops/_nested_tensor_from_tensor_list_native.h>
+#include <ATen/ops/_nested_tensor_size_native.h>
+#include <ATen/ops/_nested_tensor_softmax_with_shape_native.h>
+#include <ATen/ops/_nested_tensor_storage_offsets_native.h>
+#include <ATen/ops/_nested_tensor_strides_native.h>
+#include <ATen/ops/_nested_view_from_buffer_native.h>
+#include <ATen/ops/_nested_view_from_buffer_copy_native.h>
+#include <ATen/ops/_nested_view_from_jagged_native.h>
+#include <ATen/ops/_nested_view_from_jagged_copy_native.h>
+#include <ATen/ops/_new_zeros_with_same_feature_meta_native.h>
+#include <ATen/ops/_nnpack_available_native.h>
+#include <ATen/ops/_nnpack_spatial_convolution_native.h>
+#include <ATen/ops/_nnz_native.h>
+#include <ATen/ops/_pack_padded_sequence_native.h>
+#include <ATen/ops/_pack_padded_sequence_backward_native.h>
+#include <ATen/ops/_pad_circular_native.h>
+#include <ATen/ops/_pad_enum_native.h>
+#include <ATen/ops/_pad_packed_sequence_native.h>
+#include <ATen/ops/_pdist_backward_native.h>
+#include <ATen/ops/_pdist_forward_native.h>
+#include <ATen/ops/_pin_memory_native.h>
+#include <ATen/ops/_prelu_kernel_native.h>
+#include <ATen/ops/_prelu_kernel_backward_native.h>
+#include <ATen/ops/_print_native.h>
+#include <ATen/ops/_propagate_xla_data_native.h>
+#include <ATen/ops/_remove_batch_dim_native.h>
+#include <ATen/ops/_reshape_alias_native.h>
+#include <ATen/ops/_reshape_alias_copy_native.h>
+#include <ATen/ops/_reshape_copy_native.h>
+#include <ATen/ops/_reshape_from_tensor_native.h>
+#include <ATen/ops/_resize_output_native.h>
+#include <ATen/ops/_rowwise_prune_native.h>
+#include <ATen/ops/_sample_dirichlet_native.h>
+#include <ATen/ops/_saturate_weight_to_fp16_native.h>
+#include <ATen/ops/_scaled_dot_product_attention_math_native.h>
+#include <ATen/ops/_scaled_dot_product_cudnn_attention_native.h>
+#include <ATen/ops/_scaled_dot_product_efficient_attention_native.h>
+#include <ATen/ops/_scaled_dot_product_efficient_attention_backward_native.h>
+#include <ATen/ops/_scaled_dot_product_flash_attention_native.h>
+#include <ATen/ops/_scaled_dot_product_flash_attention_backward_native.h>
+#include <ATen/ops/_scaled_dot_product_flash_attention_for_cpu_native.h>
+#include <ATen/ops/_scaled_dot_product_flash_attention_for_cpu_backward_native.h>
+#include <ATen/ops/_scaled_mm_native.h>
+#include <ATen/ops/_segment_reduce_backward_native.h>
+#include <ATen/ops/_shape_as_tensor_native.h>
+#include <ATen/ops/_slow_conv2d_backward_native.h>
+#include <ATen/ops/_slow_conv2d_forward_native.h>
+#include <ATen/ops/_sobol_engine_draw_native.h>
+#include <ATen/ops/_sobol_engine_ff_native.h>
+#include <ATen/ops/_sobol_engine_initialize_state_native.h>
+#include <ATen/ops/_sobol_engine_scramble_native.h>
+#include <ATen/ops/_softmax_native.h>
+#include <ATen/ops/_softmax_backward_data_native.h>
+#include <ATen/ops/_sparse_addmm_native.h>
+#include <ATen/ops/_sparse_broadcast_to_native.h>
+#include <ATen/ops/_sparse_broadcast_to_copy_native.h>
+#include <ATen/ops/_sparse_bsc_tensor_unsafe_native.h>
+#include <ATen/ops/_sparse_bsr_tensor_unsafe_native.h>
+#include <ATen/ops/_sparse_compressed_tensor_unsafe_native.h>
+#include <ATen/ops/_sparse_coo_tensor_unsafe_native.h>
+#include <ATen/ops/_sparse_coo_tensor_with_dims_native.h>
+#include <ATen/ops/_sparse_coo_tensor_with_dims_and_tensors_native.h>
+#include <ATen/ops/_sparse_csc_tensor_unsafe_native.h>
+#include <ATen/ops/_sparse_csr_prod_native.h>
+#include <ATen/ops/_sparse_csr_sum_native.h>
+#include <ATen/ops/_sparse_csr_tensor_unsafe_native.h>
+#include <ATen/ops/_sparse_log_softmax_native.h>
+#include <ATen/ops/_sparse_log_softmax_backward_data_native.h>
+#include <ATen/ops/_sparse_mask_projection_native.h>
+#include <ATen/ops/_sparse_mm_native.h>
+#include <ATen/ops/_sparse_mm_reduce_impl_native.h>
+#include <ATen/ops/_sparse_mm_reduce_impl_backward_native.h>
+#include <ATen/ops/_sparse_semi_structured_linear_native.h>
+#include <ATen/ops/_sparse_softmax_native.h>
+#include <ATen/ops/_sparse_softmax_backward_data_native.h>
+#include <ATen/ops/_sparse_sparse_matmul_native.h>
+#include <ATen/ops/_sparse_sum_native.h>
+#include <ATen/ops/_sparse_sum_backward_native.h>
+#include <ATen/ops/_spdiags_native.h>
+#include <ATen/ops/_stack_native.h>
+#include <ATen/ops/_standard_gamma_native.h>
+#include <ATen/ops/_standard_gamma_grad_native.h>
+#include <ATen/ops/_test_ambiguous_defaults_native.h>
+#include <ATen/ops/_test_autograd_multiple_dispatch_native.h>
+#include <ATen/ops/_test_autograd_multiple_dispatch_view_native.h>
+#include <ATen/ops/_test_autograd_multiple_dispatch_view_copy_native.h>
+#include <ATen/ops/_test_check_tensor_native.h>
+#include <ATen/ops/_test_functorch_fallback_native.h>
+#include <ATen/ops/_test_optional_filled_intlist_native.h>
+#include <ATen/ops/_test_optional_floatlist_native.h>
+#include <ATen/ops/_test_optional_intlist_native.h>
+#include <ATen/ops/_test_parallel_materialize_native.h>
+#include <ATen/ops/_test_serialization_subcmul_native.h>
+#include <ATen/ops/_test_string_default_native.h>
+#include <ATen/ops/_test_warn_in_autograd_native.h>
+#include <ATen/ops/_thnn_differentiable_gru_cell_backward_native.h>
+#include <ATen/ops/_thnn_differentiable_lstm_cell_backward_native.h>
+#include <ATen/ops/_thnn_fused_gru_cell_native.h>
+#include <ATen/ops/_thnn_fused_gru_cell_backward_native.h>
+#include <ATen/ops/_thnn_fused_lstm_cell_native.h>
+#include <ATen/ops/_thnn_fused_lstm_cell_backward_native.h>
+#include <ATen/ops/_thnn_fused_lstm_cell_backward_impl_native.h>
+#include <ATen/ops/_to_copy_native.h>
+#include <ATen/ops/_to_cpu_native.h>
+#include <ATen/ops/_to_dense_native.h>
+#include <ATen/ops/_to_sparse_native.h>
+#include <ATen/ops/_to_sparse_bsc_native.h>
+#include <ATen/ops/_to_sparse_bsr_native.h>
+#include <ATen/ops/_to_sparse_csc_native.h>
+#include <ATen/ops/_to_sparse_csr_native.h>
+#include <ATen/ops/_to_sparse_semi_structured_native.h>
+#include <ATen/ops/_transform_bias_rescale_qkv_native.h>
+#include <ATen/ops/_transformer_encoder_layer_fwd_native.h>
+#include <ATen/ops/_trilinear_native.h>
+#include <ATen/ops/_triton_multi_head_attention_native.h>
+#include <ATen/ops/_triton_scaled_dot_attention_native.h>
+#include <ATen/ops/_unique_native.h>
+#include <ATen/ops/_unique2_native.h>
+#include <ATen/ops/_unpack_dual_native.h>
+#include <ATen/ops/_unsafe_index_native.h>
+#include <ATen/ops/_unsafe_index_put_native.h>
+#include <ATen/ops/_unsafe_view_native.h>
+#include <ATen/ops/_upsample_bicubic2d_aa_native.h>
+#include <ATen/ops/_upsample_bicubic2d_aa_backward_native.h>
+#include <ATen/ops/_upsample_bilinear2d_aa_native.h>
+#include <ATen/ops/_upsample_bilinear2d_aa_backward_native.h>
+#include <ATen/ops/_upsample_nearest_exact1d_native.h>
+#include <ATen/ops/_upsample_nearest_exact1d_backward_native.h>
+#include <ATen/ops/_upsample_nearest_exact2d_native.h>
+#include <ATen/ops/_upsample_nearest_exact2d_backward_native.h>
+#include <ATen/ops/_upsample_nearest_exact3d_native.h>
+#include <ATen/ops/_upsample_nearest_exact3d_backward_native.h>
+#include <ATen/ops/_use_cudnn_ctc_loss_native.h>
+#include <ATen/ops/_use_cudnn_rnn_flatten_weight_native.h>
+#include <ATen/ops/_validate_compressed_sparse_indices_native.h>
+#include <ATen/ops/_validate_sparse_bsc_tensor_args_native.h>
+#include <ATen/ops/_validate_sparse_bsr_tensor_args_native.h>
+#include <ATen/ops/_validate_sparse_compressed_tensor_args_native.h>
+#include <ATen/ops/_validate_sparse_coo_tensor_args_native.h>
+#include <ATen/ops/_validate_sparse_csc_tensor_args_native.h>
+#include <ATen/ops/_validate_sparse_csr_tensor_args_native.h>
+#include <ATen/ops/_values_native.h>
+#include <ATen/ops/_values_copy_native.h>
+#include <ATen/ops/_version_native.h>
+#include <ATen/ops/_weight_int4pack_mm_native.h>
+#include <ATen/ops/_weight_int8pack_mm_native.h>
+#include <ATen/ops/_weight_norm_native.h>
+#include <ATen/ops/_weight_norm_differentiable_backward_native.h>
+#include <ATen/ops/_weight_norm_interface_native.h>
+#include <ATen/ops/_weight_norm_interface_backward_native.h>
+#include <ATen/ops/abs_native.h>
+#include <ATen/ops/absolute_native.h>
+#include <ATen/ops/acos_native.h>
+#include <ATen/ops/acosh_native.h>
+#include <ATen/ops/adaptive_avg_pool1d_native.h>
+#include <ATen/ops/adaptive_avg_pool2d_native.h>
+#include <ATen/ops/adaptive_avg_pool3d_native.h>
+#include <ATen/ops/adaptive_avg_pool3d_backward_native.h>
+#include <ATen/ops/adaptive_max_pool1d_native.h>
+#include <ATen/ops/adaptive_max_pool2d_native.h>
+#include <ATen/ops/adaptive_max_pool2d_backward_native.h>
+#include <ATen/ops/adaptive_max_pool3d_native.h>
+#include <ATen/ops/adaptive_max_pool3d_backward_native.h>
+#include <ATen/ops/add_native.h>
+#include <ATen/ops/addbmm_native.h>
+#include <ATen/ops/addcdiv_native.h>
+#include <ATen/ops/addcmul_native.h>
+#include <ATen/ops/addmm_native.h>
+#include <ATen/ops/addmv_native.h>
+#include <ATen/ops/addr_native.h>
+#include <ATen/ops/adjoint_native.h>
+#include <ATen/ops/affine_grid_generator_native.h>
+#include <ATen/ops/affine_grid_generator_backward_native.h>
+#include <ATen/ops/alias_native.h>
+#include <ATen/ops/alias_copy_native.h>
+#include <ATen/ops/align_as_native.h>
+#include <ATen/ops/align_tensors_native.h>
+#include <ATen/ops/align_to_native.h>
+#include <ATen/ops/all_native.h>
+#include <ATen/ops/allclose_native.h>
+#include <ATen/ops/alpha_dropout_native.h>
+#include <ATen/ops/amax_native.h>
+#include <ATen/ops/amin_native.h>
+#include <ATen/ops/aminmax_native.h>
+#include <ATen/ops/and_native.h>
+#include <ATen/ops/angle_native.h>
+#include <ATen/ops/any_native.h>
+#include <ATen/ops/arange_native.h>
+#include <ATen/ops/arccos_native.h>
+#include <ATen/ops/arccosh_native.h>
+#include <ATen/ops/arcsin_native.h>
+#include <ATen/ops/arcsinh_native.h>
+#include <ATen/ops/arctan_native.h>
+#include <ATen/ops/arctan2_native.h>
+#include <ATen/ops/arctanh_native.h>
+#include <ATen/ops/argmax_native.h>
+#include <ATen/ops/argmin_native.h>
+#include <ATen/ops/argsort_native.h>
+#include <ATen/ops/argwhere_native.h>
+#include <ATen/ops/as_strided_native.h>
+#include <ATen/ops/as_strided_copy_native.h>
+#include <ATen/ops/as_strided_scatter_native.h>
+#include <ATen/ops/asin_native.h>
+#include <ATen/ops/asinh_native.h>
+#include <ATen/ops/atan_native.h>
+#include <ATen/ops/atan2_native.h>
+#include <ATen/ops/atanh_native.h>
+#include <ATen/ops/atleast_1d_native.h>
+#include <ATen/ops/atleast_2d_native.h>
+#include <ATen/ops/atleast_3d_native.h>
+#include <ATen/ops/avg_pool1d_native.h>
+#include <ATen/ops/avg_pool2d_native.h>
+#include <ATen/ops/avg_pool2d_backward_native.h>
+#include <ATen/ops/avg_pool3d_native.h>
+#include <ATen/ops/avg_pool3d_backward_native.h>
+#include <ATen/ops/baddbmm_native.h>
+#include <ATen/ops/bartlett_window_native.h>
+#include <ATen/ops/batch_norm_native.h>
+#include <ATen/ops/batch_norm_backward_elemt_native.h>
+#include <ATen/ops/batch_norm_backward_reduce_native.h>
+#include <ATen/ops/batch_norm_elemt_native.h>
+#include <ATen/ops/batch_norm_gather_stats_native.h>
+#include <ATen/ops/batch_norm_gather_stats_with_counts_native.h>
+#include <ATen/ops/batch_norm_stats_native.h>
+#include <ATen/ops/batch_norm_update_stats_native.h>
+#include <ATen/ops/bernoulli_native.h>
+#include <ATen/ops/bilinear_native.h>
+#include <ATen/ops/binary_cross_entropy_native.h>
+#include <ATen/ops/binary_cross_entropy_backward_native.h>
+#include <ATen/ops/binary_cross_entropy_with_logits_native.h>
+#include <ATen/ops/bincount_native.h>
+#include <ATen/ops/binomial_native.h>
+#include <ATen/ops/bitwise_and_native.h>
+#include <ATen/ops/bitwise_left_shift_native.h>
+#include <ATen/ops/bitwise_not_native.h>
+#include <ATen/ops/bitwise_or_native.h>
+#include <ATen/ops/bitwise_right_shift_native.h>
+#include <ATen/ops/bitwise_xor_native.h>
+#include <ATen/ops/blackman_window_native.h>
+#include <ATen/ops/block_diag_native.h>
+#include <ATen/ops/bmm_native.h>
+#include <ATen/ops/broadcast_tensors_native.h>
+#include <ATen/ops/broadcast_to_native.h>
+#include <ATen/ops/bucketize_native.h>
+#include <ATen/ops/can_cast_native.h>
+#include <ATen/ops/cartesian_prod_native.h>
+#include <ATen/ops/cat_native.h>
+#include <ATen/ops/cauchy_native.h>
+#include <ATen/ops/ccol_indices_native.h>
+#include <ATen/ops/ccol_indices_copy_native.h>
+#include <ATen/ops/cdist_native.h>
+#include <ATen/ops/ceil_native.h>
+#include <ATen/ops/celu_native.h>
+#include <ATen/ops/chain_matmul_native.h>
+#include <ATen/ops/chalf_native.h>
+#include <ATen/ops/channel_shuffle_native.h>
+#include <ATen/ops/cholesky_native.h>
+#include <ATen/ops/cholesky_inverse_native.h>
+#include <ATen/ops/cholesky_solve_native.h>
+#include <ATen/ops/choose_qparams_optimized_native.h>
+#include <ATen/ops/chunk_native.h>
+#include <ATen/ops/clamp_native.h>
+#include <ATen/ops/clamp_max_native.h>
+#include <ATen/ops/clamp_min_native.h>
+#include <ATen/ops/clip_native.h>
+#include <ATen/ops/clone_native.h>
+#include <ATen/ops/coalesce_native.h>
+#include <ATen/ops/col2im_native.h>
+#include <ATen/ops/col_indices_native.h>
+#include <ATen/ops/col_indices_copy_native.h>
+#include <ATen/ops/column_stack_native.h>
+#include <ATen/ops/combinations_native.h>
+#include <ATen/ops/complex_native.h>
+#include <ATen/ops/concat_native.h>
+#include <ATen/ops/concatenate_native.h>
+#include <ATen/ops/conj_native.h>
+#include <ATen/ops/conj_physical_native.h>
+#include <ATen/ops/constant_pad_nd_native.h>
+#include <ATen/ops/contiguous_native.h>
+#include <ATen/ops/conv1d_native.h>
+#include <ATen/ops/conv2d_native.h>
+#include <ATen/ops/conv3d_native.h>
+#include <ATen/ops/conv_depthwise3d_native.h>
+#include <ATen/ops/conv_tbc_native.h>
+#include <ATen/ops/conv_tbc_backward_native.h>
+#include <ATen/ops/conv_transpose1d_native.h>
+#include <ATen/ops/conv_transpose2d_native.h>
+#include <ATen/ops/conv_transpose3d_native.h>
+#include <ATen/ops/convolution_native.h>
+#include <ATen/ops/convolution_backward_native.h>
+#include <ATen/ops/convolution_backward_overrideable_native.h>
+#include <ATen/ops/convolution_overrideable_native.h>
+#include <ATen/ops/copy_native.h>
+#include <ATen/ops/copy_sparse_to_sparse_native.h>
+#include <ATen/ops/copysign_native.h>
+#include <ATen/ops/corrcoef_native.h>
+#include <ATen/ops/cos_native.h>
+#include <ATen/ops/cosh_native.h>
+#include <ATen/ops/cosine_embedding_loss_native.h>
+#include <ATen/ops/cosine_similarity_native.h>
+#include <ATen/ops/count_nonzero_native.h>
+#include <ATen/ops/cov_native.h>
+#include <ATen/ops/cross_native.h>
+#include <ATen/ops/cross_entropy_loss_native.h>
+#include <ATen/ops/crow_indices_native.h>
+#include <ATen/ops/crow_indices_copy_native.h>
+#include <ATen/ops/ctc_loss_native.h>
+#include <ATen/ops/cudnn_affine_grid_generator_native.h>
+#include <ATen/ops/cudnn_affine_grid_generator_backward_native.h>
+#include <ATen/ops/cudnn_batch_norm_native.h>
+#include <ATen/ops/cudnn_batch_norm_backward_native.h>
+#include <ATen/ops/cudnn_convolution_native.h>
+#include <ATen/ops/cudnn_convolution_add_relu_native.h>
+#include <ATen/ops/cudnn_convolution_relu_native.h>
+#include <ATen/ops/cudnn_convolution_transpose_native.h>
+#include <ATen/ops/cudnn_grid_sampler_native.h>
+#include <ATen/ops/cudnn_grid_sampler_backward_native.h>
+#include <ATen/ops/cudnn_is_acceptable_native.h>
+#include <ATen/ops/cummax_native.h>
+#include <ATen/ops/cummaxmin_backward_native.h>
+#include <ATen/ops/cummin_native.h>
+#include <ATen/ops/cumprod_native.h>
+#include <ATen/ops/cumprod_backward_native.h>
+#include <ATen/ops/cumsum_native.h>
+#include <ATen/ops/cumulative_trapezoid_native.h>
+#include <ATen/ops/data_native.h>
+#include <ATen/ops/deg2rad_native.h>
+#include <ATen/ops/dense_dim_native.h>
+#include <ATen/ops/dequantize_native.h>
+#include <ATen/ops/det_native.h>
+#include <ATen/ops/detach_native.h>
+#include <ATen/ops/detach_copy_native.h>
+#include <ATen/ops/diag_native.h>
+#include <ATen/ops/diag_embed_native.h>
+#include <ATen/ops/diagflat_native.h>
+#include <ATen/ops/diagonal_native.h>
+#include <ATen/ops/diagonal_backward_native.h>
+#include <ATen/ops/diagonal_copy_native.h>
+#include <ATen/ops/diagonal_scatter_native.h>
+#include <ATen/ops/diff_native.h>
+#include <ATen/ops/digamma_native.h>
+#include <ATen/ops/dist_native.h>
+#include <ATen/ops/div_native.h>
+#include <ATen/ops/divide_native.h>
+#include <ATen/ops/dot_native.h>
+#include <ATen/ops/dropout_native.h>
+#include <ATen/ops/dsplit_native.h>
+#include <ATen/ops/dstack_native.h>
+#include <ATen/ops/einsum_native.h>
+#include <ATen/ops/elu_native.h>
+#include <ATen/ops/elu_backward_native.h>
+#include <ATen/ops/embedding_native.h>
+#include <ATen/ops/embedding_backward_native.h>
+#include <ATen/ops/embedding_bag_native.h>
+#include <ATen/ops/embedding_dense_backward_native.h>
+#include <ATen/ops/embedding_renorm_native.h>
+#include <ATen/ops/embedding_sparse_backward_native.h>
+#include <ATen/ops/empty_native.h>
+#include <ATen/ops/empty_like_native.h>
+#include <ATen/ops/empty_permuted_native.h>
+#include <ATen/ops/empty_quantized_native.h>
+#include <ATen/ops/empty_strided_native.h>
+#include <ATen/ops/eq_native.h>
+#include <ATen/ops/equal_native.h>
+#include <ATen/ops/erf_native.h>
+#include <ATen/ops/erfc_native.h>
+#include <ATen/ops/erfinv_native.h>
+#include <ATen/ops/exp_native.h>
+#include <ATen/ops/exp2_native.h>
+#include <ATen/ops/expand_native.h>
+#include <ATen/ops/expand_as_native.h>
+#include <ATen/ops/expand_copy_native.h>
+#include <ATen/ops/expm1_native.h>
+#include <ATen/ops/exponential_native.h>
+#include <ATen/ops/eye_native.h>
+#include <ATen/ops/fake_quantize_per_channel_affine_native.h>
+#include <ATen/ops/fake_quantize_per_channel_affine_cachemask_native.h>
+#include <ATen/ops/fake_quantize_per_channel_affine_cachemask_backward_native.h>
+#include <ATen/ops/fake_quantize_per_tensor_affine_native.h>
+#include <ATen/ops/fake_quantize_per_tensor_affine_cachemask_native.h>
+#include <ATen/ops/fake_quantize_per_tensor_affine_cachemask_backward_native.h>
+#include <ATen/ops/fbgemm_linear_fp16_weight_native.h>
+#include <ATen/ops/fbgemm_linear_fp16_weight_fp32_activation_native.h>
+#include <ATen/ops/fbgemm_linear_int8_weight_native.h>
+#include <ATen/ops/fbgemm_linear_int8_weight_fp32_activation_native.h>
+#include <ATen/ops/fbgemm_linear_quantize_weight_native.h>
+#include <ATen/ops/fbgemm_pack_gemm_matrix_fp16_native.h>
+#include <ATen/ops/fbgemm_pack_quantized_matrix_native.h>
+#include <ATen/ops/feature_alpha_dropout_native.h>
+#include <ATen/ops/feature_dropout_native.h>
+#include <ATen/ops/fft_fft_native.h>
+#include <ATen/ops/fft_fft2_native.h>
+#include <ATen/ops/fft_fftfreq_native.h>
+#include <ATen/ops/fft_fftn_native.h>
+#include <ATen/ops/fft_fftshift_native.h>
+#include <ATen/ops/fft_hfft_native.h>
+#include <ATen/ops/fft_hfft2_native.h>
+#include <ATen/ops/fft_hfftn_native.h>
+#include <ATen/ops/fft_ifft_native.h>
+#include <ATen/ops/fft_ifft2_native.h>
+#include <ATen/ops/fft_ifftn_native.h>
+#include <ATen/ops/fft_ifftshift_native.h>
+#include <ATen/ops/fft_ihfft_native.h>
+#include <ATen/ops/fft_ihfft2_native.h>
+#include <ATen/ops/fft_ihfftn_native.h>
+#include <ATen/ops/fft_irfft_native.h>
+#include <ATen/ops/fft_irfft2_native.h>
+#include <ATen/ops/fft_irfftn_native.h>
+#include <ATen/ops/fft_rfft_native.h>
+#include <ATen/ops/fft_rfft2_native.h>
+#include <ATen/ops/fft_rfftfreq_native.h>
+#include <ATen/ops/fft_rfftn_native.h>
+#include <ATen/ops/fill_native.h>
+#include <ATen/ops/fill_diagonal_native.h>
+#include <ATen/ops/fix_native.h>
+#include <ATen/ops/flatten_native.h>
+#include <ATen/ops/flatten_dense_tensors_native.h>
+#include <ATen/ops/flip_native.h>
+#include <ATen/ops/fliplr_native.h>
+#include <ATen/ops/flipud_native.h>
+#include <ATen/ops/float_power_native.h>
+#include <ATen/ops/floor_native.h>
+#include <ATen/ops/floor_divide_native.h>
+#include <ATen/ops/fmax_native.h>
+#include <ATen/ops/fmin_native.h>
+#include <ATen/ops/fmod_native.h>
+#include <ATen/ops/frac_native.h>
+#include <ATen/ops/fractional_max_pool2d_native.h>
+#include <ATen/ops/fractional_max_pool2d_backward_native.h>
+#include <ATen/ops/fractional_max_pool3d_native.h>
+#include <ATen/ops/fractional_max_pool3d_backward_native.h>
+#include <ATen/ops/frexp_native.h>
+#include <ATen/ops/frobenius_norm_native.h>
+#include <ATen/ops/from_file_native.h>
+#include <ATen/ops/full_native.h>
+#include <ATen/ops/full_like_native.h>
+#include <ATen/ops/fused_moving_avg_obs_fake_quant_native.h>
+#include <ATen/ops/gather_native.h>
+#include <ATen/ops/gather_backward_native.h>
+#include <ATen/ops/gcd_native.h>
+#include <ATen/ops/ge_native.h>
+#include <ATen/ops/gelu_native.h>
+#include <ATen/ops/gelu_backward_native.h>
+#include <ATen/ops/geometric_native.h>
+#include <ATen/ops/geqrf_native.h>
+#include <ATen/ops/ger_native.h>
+#include <ATen/ops/glu_native.h>
+#include <ATen/ops/glu_backward_native.h>
+#include <ATen/ops/glu_backward_jvp_native.h>
+#include <ATen/ops/glu_jvp_native.h>
+#include <ATen/ops/gradient_native.h>
+#include <ATen/ops/greater_native.h>
+#include <ATen/ops/greater_equal_native.h>
+#include <ATen/ops/grid_sampler_native.h>
+#include <ATen/ops/grid_sampler_2d_native.h>
+#include <ATen/ops/grid_sampler_2d_backward_native.h>
+#include <ATen/ops/grid_sampler_3d_native.h>
+#include <ATen/ops/grid_sampler_3d_backward_native.h>
+#include <ATen/ops/group_norm_native.h>
+#include <ATen/ops/gru_native.h>
+#include <ATen/ops/gru_cell_native.h>
+#include <ATen/ops/gt_native.h>
+#include <ATen/ops/hamming_window_native.h>
+#include <ATen/ops/hann_window_native.h>
+#include <ATen/ops/hardshrink_native.h>
+#include <ATen/ops/hardshrink_backward_native.h>
+#include <ATen/ops/hardsigmoid_native.h>
+#include <ATen/ops/hardsigmoid_backward_native.h>
+#include <ATen/ops/hardswish_native.h>
+#include <ATen/ops/hardswish_backward_native.h>
+#include <ATen/ops/hardtanh_native.h>
+#include <ATen/ops/hardtanh_backward_native.h>
+#include <ATen/ops/heaviside_native.h>
+#include <ATen/ops/hinge_embedding_loss_native.h>
+#include <ATen/ops/histc_native.h>
+#include <ATen/ops/histogram_native.h>
+#include <ATen/ops/histogramdd_native.h>
+#include <ATen/ops/hsplit_native.h>
+#include <ATen/ops/hspmm_native.h>
+#include <ATen/ops/hstack_native.h>
+#include <ATen/ops/huber_loss_native.h>
+#include <ATen/ops/huber_loss_backward_native.h>
+#include <ATen/ops/hypot_native.h>
+#include <ATen/ops/i0_native.h>
+#include <ATen/ops/igamma_native.h>
+#include <ATen/ops/igammac_native.h>
+#include <ATen/ops/im2col_native.h>
+#include <ATen/ops/imag_native.h>
+#include <ATen/ops/index_native.h>
+#include <ATen/ops/index_add_native.h>
+#include <ATen/ops/index_copy_native.h>
+#include <ATen/ops/index_fill_native.h>
+#include <ATen/ops/index_put_native.h>
+#include <ATen/ops/index_reduce_native.h>
+#include <ATen/ops/index_select_native.h>
+#include <ATen/ops/index_select_backward_native.h>
+#include <ATen/ops/indices_native.h>
+#include <ATen/ops/indices_copy_native.h>
+#include <ATen/ops/infinitely_differentiable_gelu_backward_native.h>
+#include <ATen/ops/inner_native.h>
+#include <ATen/ops/instance_norm_native.h>
+#include <ATen/ops/int_repr_native.h>
+#include <ATen/ops/inverse_native.h>
+#include <ATen/ops/is_coalesced_native.h>
+#include <ATen/ops/is_complex_native.h>
+#include <ATen/ops/is_conj_native.h>
+#include <ATen/ops/is_distributed_native.h>
+#include <ATen/ops/is_floating_point_native.h>
+#include <ATen/ops/is_inference_native.h>
+#include <ATen/ops/is_leaf_native.h>
+#include <ATen/ops/is_neg_native.h>
+#include <ATen/ops/is_nonzero_native.h>
+#include <ATen/ops/is_pinned_native.h>
+#include <ATen/ops/is_same_size_native.h>
+#include <ATen/ops/is_set_to_native.h>
+#include <ATen/ops/is_signed_native.h>
+#include <ATen/ops/is_vulkan_available_native.h>
+#include <ATen/ops/isclose_native.h>
+#include <ATen/ops/isfinite_native.h>
+#include <ATen/ops/isin_native.h>
+#include <ATen/ops/isinf_native.h>
+#include <ATen/ops/isnan_native.h>
+#include <ATen/ops/isneginf_native.h>
+#include <ATen/ops/isposinf_native.h>
+#include <ATen/ops/isreal_native.h>
+#include <ATen/ops/istft_native.h>
+#include <ATen/ops/item_native.h>
+#include <ATen/ops/kaiser_window_native.h>
+#include <ATen/ops/kl_div_native.h>
+#include <ATen/ops/kron_native.h>
+#include <ATen/ops/kthvalue_native.h>
+#include <ATen/ops/l1_loss_native.h>
+#include <ATen/ops/layer_norm_native.h>
+#include <ATen/ops/lcm_native.h>
+#include <ATen/ops/ldexp_native.h>
+#include <ATen/ops/le_native.h>
+#include <ATen/ops/leaky_relu_native.h>
+#include <ATen/ops/leaky_relu_backward_native.h>
+#include <ATen/ops/lerp_native.h>
+#include <ATen/ops/less_native.h>
+#include <ATen/ops/less_equal_native.h>
+#include <ATen/ops/lgamma_native.h>
+#include <ATen/ops/lift_native.h>
+#include <ATen/ops/lift_fresh_native.h>
+#include <ATen/ops/lift_fresh_copy_native.h>
+#include <ATen/ops/linalg_cholesky_native.h>
+#include <ATen/ops/linalg_cholesky_ex_native.h>
+#include <ATen/ops/linalg_cond_native.h>
+#include <ATen/ops/linalg_cross_native.h>
+#include <ATen/ops/linalg_det_native.h>
+#include <ATen/ops/linalg_diagonal_native.h>
+#include <ATen/ops/linalg_eig_native.h>
+#include <ATen/ops/linalg_eigh_native.h>
+#include <ATen/ops/linalg_eigvals_native.h>
+#include <ATen/ops/linalg_eigvalsh_native.h>
+#include <ATen/ops/linalg_householder_product_native.h>
+#include <ATen/ops/linalg_inv_native.h>
+#include <ATen/ops/linalg_inv_ex_native.h>
+#include <ATen/ops/linalg_ldl_factor_native.h>
+#include <ATen/ops/linalg_ldl_factor_ex_native.h>
+#include <ATen/ops/linalg_ldl_solve_native.h>
+#include <ATen/ops/linalg_lstsq_native.h>
+#include <ATen/ops/linalg_lu_native.h>
+#include <ATen/ops/linalg_lu_factor_native.h>
+#include <ATen/ops/linalg_lu_factor_ex_native.h>
+#include <ATen/ops/linalg_lu_solve_native.h>
+#include <ATen/ops/linalg_matmul_native.h>
+#include <ATen/ops/linalg_matrix_exp_native.h>
+#include <ATen/ops/linalg_matrix_norm_native.h>
+#include <ATen/ops/linalg_matrix_power_native.h>
+#include <ATen/ops/linalg_matrix_rank_native.h>
+#include <ATen/ops/linalg_multi_dot_native.h>
+#include <ATen/ops/linalg_norm_native.h>
+#include <ATen/ops/linalg_pinv_native.h>
+#include <ATen/ops/linalg_qr_native.h>
+#include <ATen/ops/linalg_slogdet_native.h>
+#include <ATen/ops/linalg_solve_native.h>
+#include <ATen/ops/linalg_solve_ex_native.h>
+#include <ATen/ops/linalg_solve_triangular_native.h>
+#include <ATen/ops/linalg_svd_native.h>
+#include <ATen/ops/linalg_svdvals_native.h>
+#include <ATen/ops/linalg_tensorinv_native.h>
+#include <ATen/ops/linalg_tensorsolve_native.h>
+#include <ATen/ops/linalg_vander_native.h>
+#include <ATen/ops/linalg_vecdot_native.h>
+#include <ATen/ops/linalg_vector_norm_native.h>
+#include <ATen/ops/linear_native.h>
+#include <ATen/ops/linear_backward_native.h>
+#include <ATen/ops/linspace_native.h>
+#include <ATen/ops/log_native.h>
+#include <ATen/ops/log10_native.h>
+#include <ATen/ops/log1p_native.h>
+#include <ATen/ops/log2_native.h>
+#include <ATen/ops/log_normal_native.h>
+#include <ATen/ops/log_sigmoid_native.h>
+#include <ATen/ops/log_sigmoid_backward_native.h>
+#include <ATen/ops/log_sigmoid_forward_native.h>
+#include <ATen/ops/log_softmax_native.h>
+#include <ATen/ops/logaddexp_native.h>
+#include <ATen/ops/logaddexp2_native.h>
+#include <ATen/ops/logcumsumexp_native.h>
+#include <ATen/ops/logdet_native.h>
+#include <ATen/ops/logical_and_native.h>
+#include <ATen/ops/logical_not_native.h>
+#include <ATen/ops/logical_or_native.h>
+#include <ATen/ops/logical_xor_native.h>
+#include <ATen/ops/logit_native.h>
+#include <ATen/ops/logit_backward_native.h>
+#include <ATen/ops/logspace_native.h>
+#include <ATen/ops/logsumexp_native.h>
+#include <ATen/ops/lshift_native.h>
+#include <ATen/ops/lstm_native.h>
+#include <ATen/ops/lstm_cell_native.h>
+#include <ATen/ops/lstm_mps_backward_native.h>
+#include <ATen/ops/lt_native.h>
+#include <ATen/ops/lu_solve_native.h>
+#include <ATen/ops/lu_unpack_native.h>
+#include <ATen/ops/mH_native.h>
+#include <ATen/ops/mT_native.h>
+#include <ATen/ops/margin_ranking_loss_native.h>
+#include <ATen/ops/masked_fill_native.h>
+#include <ATen/ops/masked_scatter_native.h>
+#include <ATen/ops/masked_scatter_backward_native.h>
+#include <ATen/ops/masked_select_native.h>
+#include <ATen/ops/masked_select_backward_native.h>
+#include <ATen/ops/matmul_native.h>
+#include <ATen/ops/matmul_backward_native.h>
+#include <ATen/ops/matrix_H_native.h>
+#include <ATen/ops/matrix_exp_native.h>
+#include <ATen/ops/matrix_exp_backward_native.h>
+#include <ATen/ops/matrix_power_native.h>
+#include <ATen/ops/max_native.h>
+#include <ATen/ops/max_pool1d_native.h>
+#include <ATen/ops/max_pool1d_with_indices_native.h>
+#include <ATen/ops/max_pool2d_native.h>
+#include <ATen/ops/max_pool2d_backward_native.h>
+#include <ATen/ops/max_pool2d_with_indices_native.h>
+#include <ATen/ops/max_pool2d_with_indices_backward_native.h>
+#include <ATen/ops/max_pool3d_native.h>
+#include <ATen/ops/max_pool3d_with_indices_native.h>
+#include <ATen/ops/max_pool3d_with_indices_backward_native.h>
+#include <ATen/ops/max_unpool2d_native.h>
+#include <ATen/ops/max_unpool3d_native.h>
+#include <ATen/ops/maximum_native.h>
+#include <ATen/ops/mean_native.h>
+#include <ATen/ops/median_native.h>
+#include <ATen/ops/meshgrid_native.h>
+#include <ATen/ops/min_native.h>
+#include <ATen/ops/minimum_native.h>
+#include <ATen/ops/miopen_batch_norm_native.h>
+#include <ATen/ops/miopen_batch_norm_backward_native.h>
+#include <ATen/ops/miopen_convolution_native.h>
+#include <ATen/ops/miopen_convolution_add_relu_native.h>
+#include <ATen/ops/miopen_convolution_relu_native.h>
+#include <ATen/ops/miopen_convolution_transpose_native.h>
+#include <ATen/ops/miopen_depthwise_convolution_native.h>
+#include <ATen/ops/miopen_rnn_native.h>
+#include <ATen/ops/miopen_rnn_backward_native.h>
+#include <ATen/ops/mish_native.h>
+#include <ATen/ops/mish_backward_native.h>
+#include <ATen/ops/mkldnn_adaptive_avg_pool2d_native.h>
+#include <ATen/ops/mkldnn_adaptive_avg_pool2d_backward_native.h>
+#include <ATen/ops/mkldnn_convolution_native.h>
+#include <ATen/ops/mkldnn_linear_native.h>
+#include <ATen/ops/mkldnn_linear_backward_native.h>
+#include <ATen/ops/mkldnn_linear_backward_input_native.h>
+#include <ATen/ops/mkldnn_linear_backward_weights_native.h>
+#include <ATen/ops/mkldnn_max_pool2d_native.h>
+#include <ATen/ops/mkldnn_max_pool2d_backward_native.h>
+#include <ATen/ops/mkldnn_max_pool3d_native.h>
+#include <ATen/ops/mkldnn_max_pool3d_backward_native.h>
+#include <ATen/ops/mkldnn_reorder_conv2d_weight_native.h>
+#include <ATen/ops/mkldnn_reorder_conv3d_weight_native.h>
+#include <ATen/ops/mkldnn_rnn_layer_native.h>
+#include <ATen/ops/mkldnn_rnn_layer_backward_native.h>
+#include <ATen/ops/mm_native.h>
+#include <ATen/ops/mode_native.h>
+#include <ATen/ops/moveaxis_native.h>
+#include <ATen/ops/movedim_native.h>
+#include <ATen/ops/mps_convolution_backward_native.h>
+#include <ATen/ops/mps_convolution_transpose_backward_native.h>
+#include <ATen/ops/mse_loss_native.h>
+#include <ATen/ops/mse_loss_backward_native.h>
+#include <ATen/ops/msort_native.h>
+#include <ATen/ops/mul_native.h>
+#include <ATen/ops/multi_margin_loss_native.h>
+#include <ATen/ops/multi_margin_loss_backward_native.h>
+#include <ATen/ops/multilabel_margin_loss_native.h>
+#include <ATen/ops/multilabel_margin_loss_backward_native.h>
+#include <ATen/ops/multilabel_margin_loss_forward_native.h>
+#include <ATen/ops/multinomial_native.h>
+#include <ATen/ops/multiply_native.h>
+#include <ATen/ops/mv_native.h>
+#include <ATen/ops/mvlgamma_native.h>
+#include <ATen/ops/nan_to_num_native.h>
+#include <ATen/ops/nanmean_native.h>
+#include <ATen/ops/nanmedian_native.h>
+#include <ATen/ops/nanquantile_native.h>
+#include <ATen/ops/nansum_native.h>
+#include <ATen/ops/narrow_native.h>
+#include <ATen/ops/narrow_copy_native.h>
+#include <ATen/ops/native_batch_norm_native.h>
+#include <ATen/ops/native_batch_norm_backward_native.h>
+#include <ATen/ops/native_channel_shuffle_native.h>
+#include <ATen/ops/native_dropout_native.h>
+#include <ATen/ops/native_dropout_backward_native.h>
+#include <ATen/ops/native_group_norm_native.h>
+#include <ATen/ops/native_group_norm_backward_native.h>
+#include <ATen/ops/native_layer_norm_native.h>
+#include <ATen/ops/native_layer_norm_backward_native.h>
+#include <ATen/ops/native_norm_native.h>
+#include <ATen/ops/ne_native.h>
+#include <ATen/ops/neg_native.h>
+#include <ATen/ops/negative_native.h>
+#include <ATen/ops/nested_to_padded_tensor_native.h>
+#include <ATen/ops/new_empty_native.h>
+#include <ATen/ops/new_empty_strided_native.h>
+#include <ATen/ops/new_full_native.h>
+#include <ATen/ops/new_ones_native.h>
+#include <ATen/ops/new_zeros_native.h>
+#include <ATen/ops/nextafter_native.h>
+#include <ATen/ops/nll_loss_native.h>
+#include <ATen/ops/nll_loss2d_native.h>
+#include <ATen/ops/nll_loss2d_backward_native.h>
+#include <ATen/ops/nll_loss2d_forward_native.h>
+#include <ATen/ops/nll_loss_backward_native.h>
+#include <ATen/ops/nll_loss_forward_native.h>
+#include <ATen/ops/nll_loss_nd_native.h>
+#include <ATen/ops/nonzero_native.h>
+#include <ATen/ops/nonzero_numpy_native.h>
+#include <ATen/ops/nonzero_static_native.h>
+#include <ATen/ops/norm_native.h>
+#include <ATen/ops/norm_except_dim_native.h>
+#include <ATen/ops/normal_native.h>
+#include <ATen/ops/not_equal_native.h>
+#include <ATen/ops/nuclear_norm_native.h>
+#include <ATen/ops/numpy_T_native.h>
+#include <ATen/ops/one_hot_native.h>
+#include <ATen/ops/ones_native.h>
+#include <ATen/ops/ones_like_native.h>
+#include <ATen/ops/or_native.h>
+#include <ATen/ops/orgqr_native.h>
+#include <ATen/ops/ormqr_native.h>
+#include <ATen/ops/outer_native.h>
+#include <ATen/ops/output_nr_native.h>
+#include <ATen/ops/pad_native.h>
+#include <ATen/ops/pad_sequence_native.h>
+#include <ATen/ops/pairwise_distance_native.h>
+#include <ATen/ops/pdist_native.h>
+#include <ATen/ops/permute_native.h>
+#include <ATen/ops/permute_copy_native.h>
+#include <ATen/ops/pin_memory_native.h>
+#include <ATen/ops/pinverse_native.h>
+#include <ATen/ops/pixel_shuffle_native.h>
+#include <ATen/ops/pixel_unshuffle_native.h>
+#include <ATen/ops/poisson_native.h>
+#include <ATen/ops/poisson_nll_loss_native.h>
+#include <ATen/ops/polar_native.h>
+#include <ATen/ops/polygamma_native.h>
+#include <ATen/ops/positive_native.h>
+#include <ATen/ops/pow_native.h>
+#include <ATen/ops/prelu_native.h>
+#include <ATen/ops/prod_native.h>
+#include <ATen/ops/promote_types_native.h>
+#include <ATen/ops/put_native.h>
+#include <ATen/ops/q_per_channel_axis_native.h>
+#include <ATen/ops/q_per_channel_scales_native.h>
+#include <ATen/ops/q_per_channel_zero_points_native.h>
+#include <ATen/ops/q_scale_native.h>
+#include <ATen/ops/q_zero_point_native.h>
+#include <ATen/ops/qr_native.h>
+#include <ATen/ops/qscheme_native.h>
+#include <ATen/ops/quantile_native.h>
+#include <ATen/ops/quantize_per_channel_native.h>
+#include <ATen/ops/quantize_per_tensor_native.h>
+#include <ATen/ops/quantize_per_tensor_dynamic_native.h>
+#include <ATen/ops/quantized_batch_norm_native.h>
+#include <ATen/ops/quantized_gru_cell_native.h>
+#include <ATen/ops/quantized_lstm_cell_native.h>
+#include <ATen/ops/quantized_max_pool1d_native.h>
+#include <ATen/ops/quantized_max_pool2d_native.h>
+#include <ATen/ops/quantized_max_pool3d_native.h>
+#include <ATen/ops/quantized_rnn_relu_cell_native.h>
+#include <ATen/ops/quantized_rnn_tanh_cell_native.h>
+#include <ATen/ops/rad2deg_native.h>
+#include <ATen/ops/rand_native.h>
+#include <ATen/ops/rand_like_native.h>
+#include <ATen/ops/randint_native.h>
+#include <ATen/ops/randint_like_native.h>
+#include <ATen/ops/randn_native.h>
+#include <ATen/ops/randn_like_native.h>
+#include <ATen/ops/random_native.h>
+#include <ATen/ops/randperm_native.h>
+#include <ATen/ops/range_native.h>
+#include <ATen/ops/ravel_native.h>
+#include <ATen/ops/real_native.h>
+#include <ATen/ops/reciprocal_native.h>
+#include <ATen/ops/record_stream_native.h>
+#include <ATen/ops/refine_names_native.h>
+#include <ATen/ops/reflection_pad1d_native.h>
+#include <ATen/ops/reflection_pad1d_backward_native.h>
+#include <ATen/ops/reflection_pad2d_native.h>
+#include <ATen/ops/reflection_pad2d_backward_native.h>
+#include <ATen/ops/reflection_pad3d_native.h>
+#include <ATen/ops/reflection_pad3d_backward_native.h>
+#include <ATen/ops/relu_native.h>
+#include <ATen/ops/relu6_native.h>
+#include <ATen/ops/remainder_native.h>
+#include <ATen/ops/rename_native.h>
+#include <ATen/ops/renorm_native.h>
+#include <ATen/ops/repeat_native.h>
+#include <ATen/ops/repeat_interleave_native.h>
+#include <ATen/ops/replication_pad1d_native.h>
+#include <ATen/ops/replication_pad1d_backward_native.h>
+#include <ATen/ops/replication_pad2d_native.h>
+#include <ATen/ops/replication_pad2d_backward_native.h>
+#include <ATen/ops/replication_pad3d_native.h>
+#include <ATen/ops/replication_pad3d_backward_native.h>
+#include <ATen/ops/requires_grad_native.h>
+#include <ATen/ops/reshape_native.h>
+#include <ATen/ops/reshape_as_native.h>
+#include <ATen/ops/resize_native.h>
+#include <ATen/ops/resize_as_native.h>
+#include <ATen/ops/resize_as_sparse_native.h>
+#include <ATen/ops/resolve_conj_native.h>
+#include <ATen/ops/resolve_neg_native.h>
+#include <ATen/ops/result_type_native.h>
+#include <ATen/ops/retain_grad_native.h>
+#include <ATen/ops/retains_grad_native.h>
+#include <ATen/ops/rnn_relu_native.h>
+#include <ATen/ops/rnn_relu_cell_native.h>
+#include <ATen/ops/rnn_tanh_native.h>
+#include <ATen/ops/rnn_tanh_cell_native.h>
+#include <ATen/ops/roll_native.h>
+#include <ATen/ops/rot90_native.h>
+#include <ATen/ops/round_native.h>
+#include <ATen/ops/row_indices_native.h>
+#include <ATen/ops/row_indices_copy_native.h>
+#include <ATen/ops/row_stack_native.h>
+#include <ATen/ops/rrelu_native.h>
+#include <ATen/ops/rrelu_with_noise_native.h>
+#include <ATen/ops/rrelu_with_noise_backward_native.h>
+#include <ATen/ops/rshift_native.h>
+#include <ATen/ops/rsqrt_native.h>
+#include <ATen/ops/rsub_native.h>
+#include <ATen/ops/scalar_tensor_native.h>
+#include <ATen/ops/scaled_dot_product_attention_native.h>
+#include <ATen/ops/scatter_native.h>
+#include <ATen/ops/scatter_add_native.h>
+#include <ATen/ops/scatter_reduce_native.h>
+#include <ATen/ops/searchsorted_native.h>
+#include <ATen/ops/segment_reduce_native.h>
+#include <ATen/ops/select_native.h>
+#include <ATen/ops/select_backward_native.h>
+#include <ATen/ops/select_copy_native.h>
+#include <ATen/ops/select_scatter_native.h>
+#include <ATen/ops/selu_native.h>
+#include <ATen/ops/set_native.h>
+#include <ATen/ops/set_data_native.h>
+#include <ATen/ops/sgn_native.h>
+#include <ATen/ops/sigmoid_native.h>
+#include <ATen/ops/sigmoid_backward_native.h>
+#include <ATen/ops/sign_native.h>
+#include <ATen/ops/signbit_native.h>
+#include <ATen/ops/silu_native.h>
+#include <ATen/ops/silu_backward_native.h>
+#include <ATen/ops/sin_native.h>
+#include <ATen/ops/sinc_native.h>
+#include <ATen/ops/sinh_native.h>
+#include <ATen/ops/size_native.h>
+#include <ATen/ops/slice_native.h>
+#include <ATen/ops/slice_backward_native.h>
+#include <ATen/ops/slice_copy_native.h>
+#include <ATen/ops/slice_inverse_native.h>
+#include <ATen/ops/slice_scatter_native.h>
+#include <ATen/ops/slogdet_native.h>
+#include <ATen/ops/slow_conv3d_native.h>
+#include <ATen/ops/slow_conv3d_forward_native.h>
+#include <ATen/ops/slow_conv_dilated2d_native.h>
+#include <ATen/ops/slow_conv_dilated3d_native.h>
+#include <ATen/ops/slow_conv_transpose2d_native.h>
+#include <ATen/ops/slow_conv_transpose3d_native.h>
+#include <ATen/ops/smm_native.h>
+#include <ATen/ops/smooth_l1_loss_native.h>
+#include <ATen/ops/smooth_l1_loss_backward_native.h>
+#include <ATen/ops/soft_margin_loss_native.h>
+#include <ATen/ops/soft_margin_loss_backward_native.h>
+#include <ATen/ops/softmax_native.h>
+#include <ATen/ops/softplus_native.h>
+#include <ATen/ops/softplus_backward_native.h>
+#include <ATen/ops/softshrink_native.h>
+#include <ATen/ops/softshrink_backward_native.h>
+#include <ATen/ops/sort_native.h>
+#include <ATen/ops/sparse_bsc_tensor_native.h>
+#include <ATen/ops/sparse_bsr_tensor_native.h>
+#include <ATen/ops/sparse_compressed_tensor_native.h>
+#include <ATen/ops/sparse_coo_tensor_native.h>
+#include <ATen/ops/sparse_csc_tensor_native.h>
+#include <ATen/ops/sparse_csr_tensor_native.h>
+#include <ATen/ops/sparse_dim_native.h>
+#include <ATen/ops/sparse_mask_native.h>
+#include <ATen/ops/sparse_resize_native.h>
+#include <ATen/ops/sparse_resize_and_clear_native.h>
+#include <ATen/ops/sparse_sampled_addmm_native.h>
+#include <ATen/ops/special_airy_ai_native.h>
+#include <ATen/ops/special_bessel_j0_native.h>
+#include <ATen/ops/special_bessel_j1_native.h>
+#include <ATen/ops/special_bessel_y0_native.h>
+#include <ATen/ops/special_bessel_y1_native.h>
+#include <ATen/ops/special_chebyshev_polynomial_t_native.h>
+#include <ATen/ops/special_chebyshev_polynomial_u_native.h>
+#include <ATen/ops/special_chebyshev_polynomial_v_native.h>
+#include <ATen/ops/special_chebyshev_polynomial_w_native.h>
+#include <ATen/ops/special_digamma_native.h>
+#include <ATen/ops/special_entr_native.h>
+#include <ATen/ops/special_erf_native.h>
+#include <ATen/ops/special_erfc_native.h>
+#include <ATen/ops/special_erfcx_native.h>
+#include <ATen/ops/special_erfinv_native.h>
+#include <ATen/ops/special_exp2_native.h>
+#include <ATen/ops/special_expit_native.h>
+#include <ATen/ops/special_expm1_native.h>
+#include <ATen/ops/special_gammainc_native.h>
+#include <ATen/ops/special_gammaincc_native.h>
+#include <ATen/ops/special_gammaln_native.h>
+#include <ATen/ops/special_hermite_polynomial_h_native.h>
+#include <ATen/ops/special_hermite_polynomial_he_native.h>
+#include <ATen/ops/special_i0_native.h>
+#include <ATen/ops/special_i0e_native.h>
+#include <ATen/ops/special_i1_native.h>
+#include <ATen/ops/special_i1e_native.h>
+#include <ATen/ops/special_laguerre_polynomial_l_native.h>
+#include <ATen/ops/special_legendre_polynomial_p_native.h>
+#include <ATen/ops/special_log1p_native.h>
+#include <ATen/ops/special_log_ndtr_native.h>
+#include <ATen/ops/special_log_softmax_native.h>
+#include <ATen/ops/special_logit_native.h>
+#include <ATen/ops/special_logsumexp_native.h>
+#include <ATen/ops/special_modified_bessel_i0_native.h>
+#include <ATen/ops/special_modified_bessel_i1_native.h>
+#include <ATen/ops/special_modified_bessel_k0_native.h>
+#include <ATen/ops/special_modified_bessel_k1_native.h>
+#include <ATen/ops/special_multigammaln_native.h>
+#include <ATen/ops/special_ndtr_native.h>
+#include <ATen/ops/special_ndtri_native.h>
+#include <ATen/ops/special_polygamma_native.h>
+#include <ATen/ops/special_psi_native.h>
+#include <ATen/ops/special_round_native.h>
+#include <ATen/ops/special_scaled_modified_bessel_k0_native.h>
+#include <ATen/ops/special_scaled_modified_bessel_k1_native.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_t_native.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_u_native.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_v_native.h>
+#include <ATen/ops/special_shifted_chebyshev_polynomial_w_native.h>
+#include <ATen/ops/special_sinc_native.h>
+#include <ATen/ops/special_softmax_native.h>
+#include <ATen/ops/special_spherical_bessel_j0_native.h>
+#include <ATen/ops/special_xlog1py_native.h>
+#include <ATen/ops/special_xlogy_native.h>
+#include <ATen/ops/special_zeta_native.h>
+#include <ATen/ops/split_native.h>
+#include <ATen/ops/split_copy_native.h>
+#include <ATen/ops/split_with_sizes_native.h>
+#include <ATen/ops/split_with_sizes_copy_native.h>
+#include <ATen/ops/sqrt_native.h>
+#include <ATen/ops/square_native.h>
+#include <ATen/ops/squeeze_native.h>
+#include <ATen/ops/squeeze_copy_native.h>
+#include <ATen/ops/sspaddmm_native.h>
+#include <ATen/ops/stack_native.h>
+#include <ATen/ops/std_native.h>
+#include <ATen/ops/std_mean_native.h>
+#include <ATen/ops/stft_native.h>
+#include <ATen/ops/stride_native.h>
+#include <ATen/ops/sub_native.h>
+#include <ATen/ops/subtract_native.h>
+#include <ATen/ops/sum_native.h>
+#include <ATen/ops/sum_to_size_native.h>
+#include <ATen/ops/svd_native.h>
+#include <ATen/ops/swapaxes_native.h>
+#include <ATen/ops/swapdims_native.h>
+#include <ATen/ops/sym_constrain_range_native.h>
+#include <ATen/ops/sym_constrain_range_for_size_native.h>
+#include <ATen/ops/sym_numel_native.h>
+#include <ATen/ops/sym_size_native.h>
+#include <ATen/ops/sym_storage_offset_native.h>
+#include <ATen/ops/sym_stride_native.h>
+#include <ATen/ops/t_native.h>
+#include <ATen/ops/t_copy_native.h>
+#include <ATen/ops/take_native.h>
+#include <ATen/ops/take_along_dim_native.h>
+#include <ATen/ops/tan_native.h>
+#include <ATen/ops/tanh_native.h>
+#include <ATen/ops/tanh_backward_native.h>
+#include <ATen/ops/tensor_split_native.h>
+#include <ATen/ops/tensordot_native.h>
+#include <ATen/ops/thnn_conv2d_native.h>
+#include <ATen/ops/threshold_native.h>
+#include <ATen/ops/threshold_backward_native.h>
+#include <ATen/ops/tile_native.h>
+#include <ATen/ops/to_native.h>
+#include <ATen/ops/to_dense_native.h>
+#include <ATen/ops/to_dense_backward_native.h>
+#include <ATen/ops/to_mkldnn_native.h>
+#include <ATen/ops/to_mkldnn_backward_native.h>
+#include <ATen/ops/to_padded_tensor_native.h>
+#include <ATen/ops/to_sparse_native.h>
+#include <ATen/ops/to_sparse_bsc_native.h>
+#include <ATen/ops/to_sparse_bsr_native.h>
+#include <ATen/ops/to_sparse_csc_native.h>
+#include <ATen/ops/to_sparse_csr_native.h>
+#include <ATen/ops/topk_native.h>
+#include <ATen/ops/trace_native.h>
+#include <ATen/ops/trace_backward_native.h>
+#include <ATen/ops/transpose_native.h>
+#include <ATen/ops/transpose_copy_native.h>
+#include <ATen/ops/trapezoid_native.h>
+#include <ATen/ops/trapz_native.h>
+#include <ATen/ops/triangular_solve_native.h>
+#include <ATen/ops/tril_native.h>
+#include <ATen/ops/tril_indices_native.h>
+#include <ATen/ops/triplet_margin_loss_native.h>
+#include <ATen/ops/triu_native.h>
+#include <ATen/ops/triu_indices_native.h>
+#include <ATen/ops/true_divide_native.h>
+#include <ATen/ops/trunc_native.h>
+#include <ATen/ops/type_as_native.h>
+#include <ATen/ops/unbind_native.h>
+#include <ATen/ops/unbind_copy_native.h>
+#include <ATen/ops/unflatten_native.h>
+#include <ATen/ops/unflatten_dense_tensors_native.h>
+#include <ATen/ops/unfold_native.h>
+#include <ATen/ops/unfold_backward_native.h>
+#include <ATen/ops/unfold_copy_native.h>
+#include <ATen/ops/uniform_native.h>
+#include <ATen/ops/unique_consecutive_native.h>
+#include <ATen/ops/unique_dim_native.h>
+#include <ATen/ops/unique_dim_consecutive_native.h>
+#include <ATen/ops/unsafe_chunk_native.h>
+#include <ATen/ops/unsafe_split_native.h>
+#include <ATen/ops/unsafe_split_with_sizes_native.h>
+#include <ATen/ops/unsqueeze_native.h>
+#include <ATen/ops/unsqueeze_copy_native.h>
+#include <ATen/ops/upsample_bicubic2d_native.h>
+#include <ATen/ops/upsample_bicubic2d_backward_native.h>
+#include <ATen/ops/upsample_bilinear2d_native.h>
+#include <ATen/ops/upsample_bilinear2d_backward_native.h>
+#include <ATen/ops/upsample_linear1d_native.h>
+#include <ATen/ops/upsample_linear1d_backward_native.h>
+#include <ATen/ops/upsample_nearest1d_native.h>
+#include <ATen/ops/upsample_nearest1d_backward_native.h>
+#include <ATen/ops/upsample_nearest2d_native.h>
+#include <ATen/ops/upsample_nearest2d_backward_native.h>
+#include <ATen/ops/upsample_nearest3d_native.h>
+#include <ATen/ops/upsample_nearest3d_backward_native.h>
+#include <ATen/ops/upsample_trilinear3d_native.h>
+#include <ATen/ops/upsample_trilinear3d_backward_native.h>
+#include <ATen/ops/value_selecting_reduction_backward_native.h>
+#include <ATen/ops/values_native.h>
+#include <ATen/ops/values_copy_native.h>
+#include <ATen/ops/vander_native.h>
+#include <ATen/ops/var_native.h>
+#include <ATen/ops/var_mean_native.h>
+#include <ATen/ops/vdot_native.h>
+#include <ATen/ops/view_native.h>
+#include <ATen/ops/view_as_native.h>
+#include <ATen/ops/view_as_complex_native.h>
+#include <ATen/ops/view_as_complex_copy_native.h>
+#include <ATen/ops/view_as_real_native.h>
+#include <ATen/ops/view_as_real_copy_native.h>
+#include <ATen/ops/view_copy_native.h>
+#include <ATen/ops/vsplit_native.h>
+#include <ATen/ops/vstack_native.h>
+#include <ATen/ops/where_native.h>
+#include <ATen/ops/xlogy_native.h>
+#include <ATen/ops/xor_native.h>
+#include <ATen/ops/zero_native.h>
+#include <ATen/ops/zeros_native.h>
+#include <ATen/ops/zeros_like_native.h>
+
+

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/ParallelOpenMP.h

@@ -0,0 +1,54 @@
+#pragma once
+
+#include <algorithm>
+#include <atomic>
+#include <cstddef>
+#include <exception>
+
+#ifdef _OPENMP
+#define INTRA_OP_PARALLEL
+
+#include <omp.h>
+#endif
+
+#ifdef _OPENMP
+namespace at::internal {
+template <typename F>
+inline void invoke_parallel(
+    int64_t begin,
+    int64_t end,
+    int64_t grain_size,
+    const F& f) {
+  std::atomic_flag err_flag = ATOMIC_FLAG_INIT;
+  std::exception_ptr eptr;
+
+#pragma omp parallel
+  {
+    // choose number of tasks based on grain size and number of threads
+    // can't use num_threads clause due to bugs in GOMP's thread pool (See
+    // #32008)
+    int64_t num_threads = omp_get_num_threads();
+    if (grain_size > 0) {
+      num_threads = std::min(num_threads, divup((end - begin), grain_size));
+    }
+
+    int64_t tid = omp_get_thread_num();
+    int64_t chunk_size = divup((end - begin), num_threads);
+    int64_t begin_tid = begin + tid * chunk_size;
+    if (begin_tid < end) {
+      try {
+        internal::ThreadIdGuard tid_guard(tid);
+        f(begin_tid, std::min(end, chunk_size + begin_tid));
+      } catch (...) {
+        if (!err_flag.test_and_set()) {
+          eptr = std::current_exception();
+        }
+      }
+    }
+  }
+  if (eptr) {
+    std::rethrow_exception(eptr);
+  }
+}
+} // namespace at::internal
+#endif // _OPENMP

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/SmallVector.h

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

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

@@ -0,0 +1,411 @@
+#pragma once
+
+#include <ATen/SparseCsrTensorImpl.h>
+#include <ATen/SparseTensorImpl.h>
+#include <ATen/core/Tensor.h>
+
+#ifndef AT_PER_OPERATOR_HEADERS
+#include <ATen/Functions.h>
+#include <ATen/NativeFunctions.h>
+#include <ATen/Operators.h>
+#else
+#include <ATen/ops/_sparse_compressed_tensor_unsafe.h>
+#include <ATen/ops/resize_as_sparse_native.h>
+#endif
+
+#define AT_DISPATCH_ALL_SPARSE_COMPRESSED_LAYOUTS(LAYOUT, NAME, ...) \
+  [&] {                                                              \
+    const auto& the_layout = LAYOUT;                                 \
+    switch (the_layout) {                                            \
+      case kSparseCsr:                                               \
+      case kSparseCsc:                                               \
+      case kSparseBsr:                                               \
+      case kSparseBsc:                                               \
+        return __VA_ARGS__();                                        \
+      default:                                                       \
+        AT_ERROR(                                                    \
+            NAME,                                                    \
+            " expected sparse compressed tensor layout but got ",    \
+            the_layout);                                             \
+    }                                                                \
+  }()
+
+#define AT_DISPATCH_ROW_SPARSE_COMPRESSED_LAYOUTS(                \
+    LAYOUT, NAME, ROW_DIM_ACTION, COLUMN_DIM_ACTION)              \
+  [&]() {                                                         \
+    const auto& the_layout = LAYOUT;                              \
+    switch (the_layout) {                                         \
+      case kSparseCsr:                                            \
+      case kSparseBsr:                                            \
+        return (ROW_DIM_ACTION)();                                \
+      case kSparseCsc:                                            \
+      case kSparseBsc:                                            \
+        return (COLUMN_DIM_ACTION)();                             \
+      default:                                                    \
+        AT_ERROR(                                                 \
+            NAME,                                                 \
+            " expected sparse compressed tensor layout but got ", \
+            the_layout);                                          \
+    }                                                             \
+  }()
+
+#define AT_DISPATCH_PLAIN_SPARSE_COMPRESSED_LAYOUTS(              \
+    LAYOUT, NAME, NO_BLOCK_ACTION, BLOCK_ACTION)                  \
+  [&]() {                                                         \
+    const auto& the_layout = LAYOUT;                              \
+    switch (the_layout) {                                         \
+      case kSparseCsr:                                            \
+      case kSparseCsc:                                            \
+        return (NO_BLOCK_ACTION)();                               \
+      case kSparseBsr:                                            \
+      case kSparseBsc:                                            \
+        return (BLOCK_ACTION)();                                  \
+      default:                                                    \
+        AT_ERROR(                                                 \
+            NAME,                                                 \
+            " expected sparse compressed tensor layout but got ", \
+            the_layout);                                          \
+    }                                                             \
+  }()
+
+#define AT_DISPATCH_SPARSE_ROW_COMPRESSED_LAYOUTS(                    \
+    LAYOUT, NAME, ROW_DIM_ACTION)                                     \
+  [&]() {                                                             \
+    const auto& the_layout = LAYOUT;                                  \
+    switch (the_layout) {                                             \
+      case kSparseCsr:                                                \
+      case kSparseBsr:                                                \
+        return (ROW_DIM_ACTION)();                                    \
+      default:                                                        \
+        AT_ERROR(                                                     \
+            NAME,                                                     \
+            " expected sparse row compressed tensor layout but got ", \
+            the_layout);                                              \
+    }                                                                 \
+  }()
+
+#define AT_DISPATCH_SPARSE_COL_COMPRESSED_LAYOUTS(                       \
+    LAYOUT, NAME, COL_DIM_ACTION)                                        \
+  [&]() {                                                                \
+    const auto& the_layout = LAYOUT;                                     \
+    switch (the_layout) {                                                \
+      case kSparseCsc:                                                   \
+      case kSparseBsc:                                                   \
+        return (COL_DIM_ACTION)();                                       \
+      default:                                                           \
+        AT_ERROR(                                                        \
+            NAME,                                                        \
+            " expected sparse column compressed tensor layout but got ", \
+            the_layout);                                                 \
+    }                                                                    \
+  }()
+
+#define AT_DISPATCH_SPARSE_COMPRESSED_NONBLOCK_LAYOUTS(LAYOUT, NAME, ACTION)  \
+  [&]() {                                                                     \
+    const auto& the_layout = LAYOUT;                                          \
+    switch (the_layout) {                                                     \
+      case kSparseCsr:                                                        \
+      case kSparseCsc:                                                        \
+        return (ACTION)();                                                    \
+      default:                                                                \
+        AT_ERROR(                                                             \
+            NAME,                                                             \
+            " expected sparse compressed (non-block) tensor layout but got ", \
+            the_layout);                                                      \
+    }                                                                         \
+  }()
+
+#define AT_DISPATCH_SPARSE_COMPRESSED_BLOCK_LAYOUTS(LAYOUT, NAME, ACTION) \
+  [&]() {                                                                 \
+    const auto& the_layout = LAYOUT;                                      \
+    switch (the_layout) {                                                 \
+      case kSparseBsr:                                                    \
+      case kSparseBsc:                                                    \
+        return (ACTION)();                                                \
+      default:                                                            \
+        AT_ERROR(                                                         \
+            NAME,                                                         \
+            " expected sparse compressed block tensor layout but got ",   \
+            the_layout);                                                  \
+    }                                                                     \
+  }()
+
+#define AT_DISPATCH_SPARSE_VALUE_TYPES(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(                                   \
+      TYPE,                                             \
+      NAME,                                             \
+      AT_DISPATCH_CASE_ALL_TYPES_AND_COMPLEX_AND4(      \
+          kComplexHalf, kHalf, kBool, kBFloat16, __VA_ARGS__))
+
+namespace at::sparse_csr {
+
+using SparseCsrTensor = Tensor;
+
+inline bool is_sparse_compressed(const Layout& layout) {
+  switch (layout) {
+    case kSparseCsr:
+    case kSparseCsc:
+    case kSparseBsr:
+    case kSparseBsc:
+      return true;
+    default:;
+  }
+  return false;
+}
+
+inline bool is_sparse_compressed(const Tensor& self) {
+  return is_sparse_compressed(self.layout());
+}
+
+inline SparseCsrTensorImpl* get_sparse_csr_impl(const SparseCsrTensor& self) {
+  AT_DISPATCH_ALL_SPARSE_COMPRESSED_LAYOUTS(
+      self.layout(), "get_sparse_csr_impl", [&] {});
+  return static_cast<SparseCsrTensorImpl*>(self.unsafeGetTensorImpl());
+}
+
+inline std::string layoutToString(
+    Layout layout,
+    bool upper = false,
+    bool lower = false) {
+  switch (layout) {
+    case kSparseCsr:
+      return (upper ? "CSR" : (lower ? "csr" : "Csr"));
+    case kSparseCsc:
+      return (upper ? "CSC" : (lower ? "csc" : "Csc"));
+    case kSparseBsr:
+      return (upper ? "BSR" : (lower ? "bsr" : "Bsr"));
+    case kSparseBsc:
+      return (upper ? "BSC" : (lower ? "bsc" : "Bsc"));
+    default:
+      TORCH_CHECK(false, "Not a sparse compressed layout:", layout);
+      return "";
+  }
+}
+
+inline bool isCompressedRow(Layout layout) {
+  return AT_DISPATCH_ROW_SPARSE_COMPRESSED_LAYOUTS(
+      layout, "isCompressedRow", [&] { return true; }, [&] { return false; });
+}
+
+inline bool isCompressedColumn(Layout layout) {
+  return AT_DISPATCH_ROW_SPARSE_COMPRESSED_LAYOUTS(
+      layout,
+      "isCompressedColumn",
+      [&] { return false; },
+      [&] { return true; });
+}
+
+inline std::string compressedIndicesName(Layout layout) {
+  return AT_DISPATCH_ROW_SPARSE_COMPRESSED_LAYOUTS(
+      layout,
+      "compressedIndicesName",
+      [&] { return "crow_indices"; },
+      [&] { return "ccol_indices"; });
+}
+
+inline std::string plainIndicesName(Layout layout) {
+  return AT_DISPATCH_ROW_SPARSE_COMPRESSED_LAYOUTS(
+      layout,
+      "plainIndicesName",
+      [&] { return "col_indices"; },
+      [&] { return "row_indices"; });
+}
+
+inline std::string compressedDimName(Layout layout) {
+  switch (layout) {
+    case kSparseCsr:
+      return "row";
+    case kSparseCsc:
+      return "column";
+    case kSparseBsr:
+      return "row block";
+    case kSparseBsc:
+      return "column block";
+    default:
+      TORCH_CHECK(false, "Not a sparse compressed layout:", layout);
+      return "";
+  }
+}
+
+inline std::string plainDimName(Layout layout) {
+  switch (layout) {
+    case kSparseCsr:
+      return "column";
+    case kSparseCsc:
+      return "row";
+    case kSparseBsr:
+      return "column block";
+    case kSparseBsc:
+      return "row block";
+    default:
+      TORCH_CHECK(false, "Not a sparse compressed layout:", layout);
+      return "";
+  }
+}
+
+inline size_t rowDimension(Layout layout, IntArrayRef size) {
+  return size.size() - (isCompressedRow(layout) ? 2 : 1);
+}
+
+inline size_t columnDimension(Layout layout, IntArrayRef size) {
+  return size.size() - (isCompressedColumn(layout) ? 2 : 1);
+}
+
+inline size_t compressedDimension(
+    Layout layout,
+    IntArrayRef size,
+    size_t dense_ndim = 0) {
+  return size.size() - dense_ndim - (isCompressedRow(layout) ? 2 : 1);
+}
+
+inline size_t plainDimension(
+    Layout layout,
+    IntArrayRef size,
+    size_t dense_ndim = 0) {
+  return size.size() - dense_ndim - (isCompressedRow(layout) ? 1 : 2);
+}
+
+inline int64_t numBatchDimensions(Tensor const& self) {
+  return AT_DISPATCH_ROW_SPARSE_COMPRESSED_LAYOUTS(
+      self.layout(),
+      "numBatchDimensions",
+      [&self] { return self.crow_indices().dim() - 1; },
+      [&self] { return self.ccol_indices().dim() - 1; });
+}
+
+inline std::pair<Tensor, Tensor> getCompressedPlainIndices(Tensor const& self) {
+  return AT_DISPATCH_ROW_SPARSE_COMPRESSED_LAYOUTS(
+      self.layout(),
+      "getCompressedPlainIndices",
+      [&self] {
+        return std::make_pair(self.crow_indices(), self.col_indices());
+      },
+      [&self] {
+        return std::make_pair(self.ccol_indices(), self.row_indices());
+      });
+}
+
+inline Layout flip_compressed_layout(Layout layout) {
+  switch (layout) {
+    case kSparseCsr:
+      return kSparseCsc;
+    case kSparseCsc:
+      return kSparseCsr;
+    case kSparseBsr:
+      return kSparseBsc;
+    case kSparseBsc:
+      return kSparseBsr;
+    default:
+      TORCH_CHECK(false, "Not a sparse compressed layout:", layout);
+      return kSparseCsr;
+  }
+}
+
+inline DimVector getBlockSize(Tensor const& self) {
+  int64_t n_batch = numBatchDimensions(self);
+  return at::DimVector(self.values().sizes().slice(n_batch + 1, 2));
+}
+
+inline at::OptionalArray<at::SymInt> getSymIntBlockSize(Tensor const& self) {
+  if (self.layout() == at::kSparseBsr || self.layout() == at::kSparseBsc) {
+    int64_t n_batch = numBatchDimensions(self);
+    return self.values().sym_sizes().slice(n_batch + 1, 2).vec();
+  } else {
+    return {};
+  }
+}
+
+template <typename binary_op_t, typename binary_op_out_t>
+inline bool only_sparse_compressed_binary_op_trivial_cases(
+    const Tensor& self,
+    const Tensor& other,
+    const Scalar& alpha,
+    Tensor& out,
+    const binary_op_t& binary_op,
+    const binary_op_out_t& binary_op_out) {
+  // Only sparse compressed! Just like the name says :)
+  TORCH_INTERNAL_ASSERT(at::sparse_csr::is_sparse_compressed(self));
+  TORCH_INTERNAL_ASSERT(at::sparse_csr::is_sparse_compressed(other));
+  TORCH_INTERNAL_ASSERT(at::sparse_csr::is_sparse_compressed(out));
+
+  // Bypass BLAS if there are matches in (self, other, out)
+  if (self.is_same(out) && self.is_same(other)) {
+    binary_op_out(self.values(), other.values(), alpha);
+    return true;
+  }
+  if (self.is_same(other)) {
+    auto [compressed_indices, plain_indices] =
+        at::sparse_csr::getCompressedPlainIndices(self);
+    static_cast<SparseCsrTensorImpl*>(out.unsafeGetTensorImpl())
+        ->set_member_tensors(
+            compressed_indices,
+            plain_indices,
+            binary_op(self.values(), other.values(), alpha),
+            self.sizes());
+    return true;
+  }
+  return false;
+}
+
+inline bool only_sparse_compressed_add_trivial_cases(
+    const Tensor& self,
+    const Tensor& other,
+    const Scalar& alpha,
+    Tensor& out) {
+  return only_sparse_compressed_binary_op_trivial_cases(
+      self,
+      other,
+      alpha,
+      out,
+      [](const Tensor& v1, const Tensor& v2, const Scalar& alpha) {
+        return v1.add(v2, alpha);
+      },
+      [](const Tensor& v1, const Tensor& v2, const Scalar& alpha) {
+        return v1.add_(v2, alpha);
+      });
+}
+
+inline Tensor to_type(const Tensor& input, ScalarType dtype) {
+  auto [compressed_indices, plain_indices] =
+      at::sparse_csr::getCompressedPlainIndices(input);
+  return at::_sparse_compressed_tensor_unsafe(
+      compressed_indices,
+      plain_indices,
+      std::move(input.values()).to(dtype),
+      input.sizes(),
+      dtype,
+      input.layout(),
+      input.device(),
+      input.options().pinned_memory_opt());
+}
+
+template <typename acc_t, typename scalar_t>
+inline std::tuple<Tensor, Tensor> create_acc_buffer(
+    TensorOptions option,
+    ScalarType type,
+    int64_t nnz = -1) {
+  Tensor new_values, new_values_acc;
+  constexpr bool need_acc = !std::is_same_v<scalar_t, acc_t>;
+  bool is_integral = at::isIntegralType(type, /*includeBool=*/true);
+  if constexpr (need_acc) {
+    auto acc_dtype = CppTypeToScalarType<acc_t>::value;
+    new_values_acc = at::empty({}, option.dtype(acc_dtype));
+    new_values = is_integral ? new_values_acc : at::empty({}, option);
+  } else {
+    new_values = new_values_acc = at::empty({}, option);
+  }
+  if (nnz != -1) {
+    return std::make_tuple(
+        new_values.resize_(nnz), new_values_acc.resize_(nnz));
+  } else {
+    return std::make_tuple(new_values, new_values_acc);
+  }
+}
+
+inline void copy_from_acc_buffer(Tensor& new_values, Tensor& new_values_acc) {
+  if (!new_values_acc.is_same(new_values)) {
+    new_values.copy_(new_values_acc);
+  }
+}
+
+} // namespace at::sparse_csr

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/TensorGeometry.h

@@ -0,0 +1,144 @@
+#pragma once
+
+#include <ATen/core/TensorBase.h>
+#include <c10/core/WrapDimMinimal.h>
+
+namespace at {
+
+// Return if the tensor geometry represented by `sizes` and `strides` is
+// contiguous Although we cache is_contiguous in tensor now, this is till useful
+// because it allows checking if a particular geometry is contiguous without
+// explicitly constructing a tensor, e.g., when you want to choose a kernel
+// strategy based on whether a subgeometry is contiguous.
+TORCH_API bool geometry_is_contiguous(IntArrayRef sizes, IntArrayRef strides);
+
+struct TORCH_API TensorGeometry {
+  TensorGeometry() = default;
+
+  explicit TensorGeometry(c10::SymIntArrayRef sizes)
+      : sizes_(sizes.vec()),
+        strides_(sizes.size()),
+        has_symbolic_sizes_strides_(
+            !c10::asIntArrayRefSlowOpt(sizes).has_value()) {
+    int64_t dim = static_cast<int64_t>(sizes.size());
+    c10::SymInt expected_stride = 1;
+    for (int64_t i = dim - 1; i >= 0; i--) {
+      strides_[i] = expected_stride;
+      expected_stride *= sizes_[i];
+    }
+    numel_ = expected_stride;
+  }
+
+  explicit TensorGeometry(const TensorBase& t)
+      : sizes_(t.sym_sizes().vec()),
+        strides_(t.sym_strides().vec()),
+        storage_offset_(t.sym_storage_offset()),
+        numel_(t.sym_numel()),
+        has_symbolic_sizes_strides_(
+            t.unsafeGetTensorImpl()->has_symbolic_sizes_strides()) {}
+
+  // true if the tensor is contiguous
+  bool is_contiguous() const;
+
+  int64_t dim() const {
+    return static_cast<int64_t>(sizes_.size());
+  }
+
+  int64_t size(int64_t dim) const {
+    TORCH_INTERNAL_ASSERT(!has_symbolic_sizes_strides_);
+    dim = c10::maybe_wrap_dim(dim, this->dim());
+    return sizes_.at(static_cast<size_t>(dim)).as_int_unchecked();
+  }
+  c10::IntArrayRef sizes() const {
+    TORCH_INTERNAL_ASSERT(!has_symbolic_sizes_strides_);
+    return c10::asIntArrayRefUnchecked(sizes_);
+  }
+  int64_t stride(int64_t dim) const {
+    TORCH_INTERNAL_ASSERT(!has_symbolic_sizes_strides_);
+    dim = c10::maybe_wrap_dim(dim, this->dim());
+    return strides_.at(static_cast<size_t>(dim)).as_int_unchecked();
+  }
+  c10::IntArrayRef strides() const {
+    TORCH_INTERNAL_ASSERT(!has_symbolic_sizes_strides_);
+    return c10::asIntArrayRefUnchecked(strides_);
+  }
+  int64_t storage_offset() const {
+    TORCH_INTERNAL_ASSERT(!has_symbolic_sizes_strides_);
+    return storage_offset_.as_int_unchecked();
+  }
+  int64_t numel() const {
+    TORCH_INTERNAL_ASSERT(!has_symbolic_sizes_strides_);
+    return numel_.as_int_unchecked();
+  }
+
+  c10::SymInt sym_size(int64_t dim) const {
+    dim = c10::maybe_wrap_dim(dim, this->dim());
+    return sizes_.at(static_cast<size_t>(dim));
+  }
+  c10::SymIntArrayRef sym_sizes() const {
+    return sizes_;
+  }
+  c10::SymInt sym_stride(int64_t dim) const {
+    dim = c10::maybe_wrap_dim(dim, this->dim());
+    return strides_.at(static_cast<size_t>(dim));
+  }
+  c10::SymIntArrayRef sym_strides() const {
+    return strides_;
+  }
+  c10::SymInt sym_storage_offset() const {
+    return storage_offset_;
+  }
+  c10::SymInt sym_numel() const {
+    return numel_;
+  }
+
+  TensorGeometry transpose(int64_t dim0, int64_t dim1) {
+    TensorGeometry r = *this; // copy
+    TORCH_CHECK(
+        dim0 < dim(),
+        "transpose: dim0=",
+        dim0,
+        " out of range (dim=",
+        dim(),
+        ")")
+    TORCH_CHECK(
+        dim1 < dim(),
+        "transpose: dim1=",
+        dim1,
+        " out of range (dim=",
+        dim(),
+        ")")
+    std::swap(r.sizes_[dim0], r.sizes_[dim1]);
+    std::swap(r.strides_[dim0], r.strides_[dim1]);
+    return r;
+  }
+
+  std::vector<c10::SymInt>& mutable_sizes() {
+    return sizes_;
+  }
+  std::vector<c10::SymInt>& mutable_strides() {
+    return strides_;
+  }
+  c10::SymInt& mutable_storage_offset() {
+    return storage_offset_;
+  }
+  void recompute() {
+    // recalculate numel after a change
+    c10::SymInt numel = 1;
+    for (const auto& i : sizes_) {
+      numel = numel * i;
+    }
+    numel_ = std::move(numel);
+    has_symbolic_sizes_strides_ =
+        !c10::asIntArrayRefSlowOpt(sizes_).has_value();
+  }
+
+ private:
+  std::vector<c10::SymInt> sizes_;
+  std::vector<c10::SymInt> strides_;
+  c10::SymInt storage_offset_;
+  c10::SymInt numel_;
+  bool has_symbolic_sizes_strides_{false};
+};
+
+} // namespace at

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/TracerMode.h

@@ -0,0 +1,132 @@
+#pragma once
+
+#include <c10/core/impl/LocalDispatchKeySet.h>
+#include <c10/macros/Export.h>
+#include <c10/macros/Macros.h>
+
+// NOTE [Tracing Mode Switches]
+//
+// Historically, tracing function was controlled by two switches:
+//
+// - `AutoDispatchBelowADInplaceOrView` guard
+//
+//    Tracing function used to be script-generated inside `VariableType_*.cpp`
+//    kernels, sharing the same `Autograd` dispatch key with autograd function.
+//    Therefore, before tracing function was moved out of VariableType,
+//    `AutoDispatchBelowADInplaceOrView` guard can also disable tracing as a
+//    side effect of disabling `Autograd` dispatching.
+//
+// - `setTracingState()` API in `torch/csrc/jit/frontend/tracer.h`
+//
+//    It stores tracing data in a `TracingState` object in TLS. If the
+//    `TracingState` object in TLS is `null`, then tracing is paused.
+//
+//    The `TracingState` object is created in `tracer::trace()` - the main
+//    entrance of tracing function. It's temporarily set to `null` inside
+//    generated VariableType (now TraceType) to bypass tracing for intermediate
+//    ops (ops being called by other ops). After the intermediate op call
+//    finishes it's set back to the original `TracingState` object.
+//
+//    The `TracingState` obect in TLS can also be read/written via its Python
+//    binding in `python_tracer.cpp`, and `get/setTracingState()` C++ APIs,
+//    which are also exposed as `TORCH_API`.
+//
+// Two new switches were introduced since tracing function was moved out of
+// VariableType:
+//
+// - `tracer::impl::set_dispatch_enabled()` API
+//
+//    Unlike the special `Autograd` dispatch key which is included in dispatch
+//    key set by default, `Tracer` dispatch key is off by default. The
+//    dispatching switch can be toggled via this new API.
+//
+// - `tracer::impl::NoTracerDispatchMode` guard
+//
+//    It's used to cover the old semantics of `AutoDispatchBelowADInplaceOrView`
+//    after tracing was moved out of VariableType.
+//
+// Before tracing function was moved out of VariableType, tracing was enabled
+// when the following conditions are satisfied:
+//
+//    1) `TracingState` object in TLS != null;
+//       - Either inside the execution scope of `tracer::trace()`, or
+//       - Eagerly called `setTracingState()` with non-null object.
+//    2) Not inside `AutoDispatchBelowADInplaceOrView` scope;
+//
+// After:
+//
+//    1) `TracingState` object in TLS != null;
+//    2) Has called `tracer::impl::set_dispatch_enabled(true)`;
+//    3) Not inside `tracer::impl::NonDispatchGuard` scope;
+//
+// [TODOs]
+//
+// - `setTracingState()` v.s. `tracer::impl::set_dispatch_enabled()`
+//
+//   Currently `set_dispatch_enabled()` is set/unset inside `setTracingState()`
+//   to keep the semantics exactly the same as before - it's confusing to keep
+//   both switches, though. We should consider simplifying/limiting the exposed
+//   `setTracingState()` Python/C++ APIs (and other APIs calling it) so that
+//   these two can be unified.
+//
+// - `AutoDispatchBelowADInplaceOrView` v.s.
+// `tracer::impl::NoTracerDispatchMode`
+//
+//   We don't need to always set both guards together to keep semantics
+//   unchanged. For the follow use cases of `AutoDispatchBelowADInplaceOrView`
+//   we don't need set the new tracer guard:
+//
+//   * Script-generated VariableType kernels. The guard is not necessary as
+//     tracing is already disabled explicitly by `setTracingState(null)` in
+//     generated TraceType kernels - we could keep it as is or use the new guard
+//     instead.
+//
+//   * Custom ops. Will be handled by fallback kernel for `Tracer`.
+//
+//   * Functions that are not likely to be called in tracing context (no python
+//     binding / not an operator), e.g.: all mobile forward() wrappers, test
+//     binaries, and etc.
+//
+//   * Where new threads are spawned, e.g.: ATen/native/ConvolutionMM2d.cpp.
+//     It's not necessary as tracing is off by default.
+//
+//   For the rest of cases we might need have both:
+//
+//   * Functions that might be reachable from eager mode python (especially
+//     factory methods), e.g.:
+//     `internal_new_from_data()` in `torch/csrc/utils/tensor_new.cpp`.
+//     Without the new guard it will add `aten::empty` to the traced graph.
+//
+//   * Some manually maintained functions, e.g.:
+//     `torch/csrc/autograd/VariableTypeManual.cpp`.
+//     Set the new guard if it's not obvious whether `setTracingState(null)`
+//     has been called before it reaches the `AutoDispatchBelowADInplaceOrView`
+//     guard.
+//
+//   We might need tweak the usage of the new guard to optimize/fix things.
+//   It should only affect the correctness of tracing function, because the
+//   guard is essentially no-op when the master `setTracingState()` switch is
+//   off.
+
+// TODO: move this from `at::` to `jit::torch::` after
+// `aten/src/ATen/cpp_custom_type_hack.h` is removed.
+
+namespace at::tracer::impl {
+
+static inline bool is_dispatch_enabled() {
+  return c10::impl::tls_is_dispatch_key_included(at::DispatchKey::Tracer) &&
+      !c10::impl::tls_is_dispatch_key_excluded(at::DispatchKey::Tracer);
+}
+
+static inline void set_dispatch_enabled(bool enabled) {
+  TORCH_INTERNAL_ASSERT(
+      !c10::impl::tls_is_dispatch_key_excluded(at::DispatchKey::Tracer),
+      "Cannot enable tracing within the scope of NoTracerDispatchMode!");
+  c10::impl::tls_set_dispatch_key_included(at::DispatchKey::Tracer, enabled);
+}
+
+struct NoTracerDispatchMode {
+  c10::impl::ExcludeDispatchKeyGuard guard_{at::DispatchKey::Tracer};
+};
+
+} // namespace at::tracer::impl

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/autocast_mode.h

@@ -0,0 +1,647 @@
+#pragma once
+
+#include <ATen/ATen.h>
+#include <ATen/NativeFunctions.h>
+#include <ATen/Operators.h>
+#include <torch/library.h>
+
+#include <c10/core/impl/LocalDispatchKeySet.h>
+#include <c10/util/intrusive_ptr.h>
+
+namespace at::autocast {
+
+TORCH_API bool is_enabled();
+TORCH_API void set_enabled(bool enabled);
+TORCH_API void clear_cache();
+TORCH_API int increment_nesting();
+TORCH_API int decrement_nesting();
+TORCH_API bool is_cpu_enabled();
+TORCH_API void set_cpu_enabled(bool enabled);
+TORCH_API at::ScalarType get_autocast_gpu_dtype();
+TORCH_API at::ScalarType get_autocast_cpu_dtype();
+TORCH_API void set_autocast_gpu_dtype(at::ScalarType dtype);
+TORCH_API void set_autocast_cpu_dtype(at::ScalarType dtype);
+TORCH_API bool is_xpu_enabled();
+TORCH_API void set_xpu_enabled(bool enabled);
+TORCH_API at::ScalarType get_autocast_xpu_dtype();
+TORCH_API void set_autocast_xpu_dtype(at::ScalarType dtype);
+TORCH_API bool is_ipu_enabled();
+TORCH_API void set_ipu_enabled(bool enabled);
+TORCH_API at::ScalarType get_autocast_ipu_dtype();
+TORCH_API void set_autocast_ipu_dtype(at::ScalarType dtype);
+TORCH_API bool is_hpu_enabled();
+TORCH_API void set_hpu_enabled(bool enabled);
+TORCH_API at::ScalarType get_autocast_hpu_dtype();
+TORCH_API void set_autocast_hpu_dtype(at::ScalarType dtype);
+TORCH_API bool is_xla_enabled();
+TORCH_API void set_xla_enabled(bool enabled);
+TORCH_API at::ScalarType get_autocast_xla_dtype();
+TORCH_API void set_autocast_xla_dtype(at::ScalarType dtype);
+TORCH_API bool is_privateuseone_enabled();
+TORCH_API void set_privateuseone_enabled(bool enabled);
+TORCH_API at::ScalarType get_autocast_privateuseone_dtype();
+TORCH_API void set_autocast_privateuseone_dtype(at::ScalarType dtype);
+TORCH_API bool is_autocast_cache_enabled();
+TORCH_API void set_autocast_cache_enabled(bool enabled);
+
+namespace {
+inline bool is_autocast_eligible(
+    const Tensor& tensor,
+    c10::DeviceType device_type) {
+  switch (device_type) {
+    case c10::DeviceType::CUDA:
+      return (tensor.is_cuda() || tensor.is_xla()) &&
+          tensor.is_floating_point();
+    case c10::DeviceType::CPU:
+      return (tensor.is_cpu() || tensor.is_mkldnn()) &&
+          tensor.is_floating_point();
+    case c10::DeviceType::XPU:
+      return tensor.is_xpu() && tensor.is_floating_point();
+    case c10::DeviceType::IPU:
+      return tensor.is_ipu() && tensor.is_floating_point();
+    case c10::DeviceType::HPU:
+      return tensor.is_hpu() && tensor.is_floating_point();
+    case c10::DeviceType::XLA:
+      return tensor.is_xla() && tensor.is_floating_point();
+    case c10::DeviceType::PrivateUse1:
+      return tensor.is_privateuseone() && tensor.is_floating_point();
+    default:
+      return false;
+  }
+}
+} // namespace
+
+inline DispatchKey get_autocast_dispatch_key_from_device_type(
+    c10::DeviceType device_type) {
+  switch (device_type) {
+    case c10::DeviceType::CUDA:
+      return DispatchKey::Autocast;
+    case c10::DeviceType::CPU:
+      return DispatchKey::AutocastCPU;
+    case c10::DeviceType::XPU:
+      return DispatchKey::AutocastXPU;
+    case c10::DeviceType::IPU:
+      return DispatchKey::AutocastIPU;
+    case c10::DeviceType::HPU:
+      return DispatchKey::AutocastHPU;
+    case c10::DeviceType::XLA:
+      return DispatchKey::AutocastXLA;
+    case c10::DeviceType::PrivateUse1:
+      return DispatchKey::AutocastPrivateUse1;
+    default:
+      throw std::runtime_error(
+          "unknown device type for autocast in get_autocast_dispatch_key_from_device_type");
+  }
+}
+
+inline at::ScalarType get_lower_precision_fp_from_device_type(
+    c10::DeviceType device_type) {
+  switch (device_type) {
+    case c10::DeviceType::CUDA:
+      return get_autocast_gpu_dtype();
+    case c10::DeviceType::CPU:
+      return get_autocast_cpu_dtype();
+    case c10::DeviceType::XPU:
+      return get_autocast_xpu_dtype();
+    case c10::DeviceType::IPU:
+      return get_autocast_ipu_dtype();
+    case c10::DeviceType::HPU:
+      return get_autocast_hpu_dtype();
+    case c10::DeviceType::XLA:
+      return get_autocast_xla_dtype();
+    case c10::DeviceType::PrivateUse1:
+      return get_autocast_privateuseone_dtype();
+    default:
+      throw std::runtime_error(
+          "unknown device type for autocast in get_lower_precision_fp_from_device_type");
+  }
+}
+
+/********************************************************************
+Logic to extract the promote type from any Tensor or TensorList args.
+********************************************************************/
+
+// Overload to catch Tensor args.
+// If nextArg is floating-point, compare its scalar_type with our
+// current best guess for the promote type, and update if necessary.
+inline at::ScalarType prioritize(
+    at::ScalarType current,
+    const Tensor& nextArg,
+    c10::DeviceType device_type = c10::DeviceType::CUDA) {
+  if (current == at::kDouble) {
+    AT_ERROR("promote type is double in at::autocast::prioritize");
+    return current;
+  }
+  at::ScalarType lower_precision_fp =
+      get_lower_precision_fp_from_device_type(device_type);
+  if (is_autocast_eligible(nextArg, device_type)) {
+    auto next = nextArg.scalar_type();
+    if (next == at::kDouble) {
+      return current; // ignores double tensors
+    } else if (current == at::kFloat || next == at::kFloat) {
+      return at::kFloat; // prioritizes float over lower_precision_fp
+    } else if (current == lower_precision_fp && next == lower_precision_fp) {
+      return lower_precision_fp;
+    } else {
+      AT_ERROR("Unexpected floating ScalarType in at::autocast::prioritize");
+      return current;
+    }
+  } else {
+    return current;
+  }
+}
+
+// Overload to catch TensorList args (for e.g. cat, stack).
+// Reuses the overload above to process each Tensor in the list.
+inline at::ScalarType prioritize(
+    at::ScalarType current,
+    const TensorList& list,
+    c10::DeviceType device_type = c10::DeviceType::CUDA) {
+  for (const auto& tensor : list) {
+    current = prioritize(current, tensor, device_type);
+  }
+  return current;
+}
+
+inline at::ScalarType prioritize(
+    at::ScalarType current,
+    const ITensorListRef& list,
+    c10::DeviceType device_type = c10::DeviceType::CUDA) {
+  for (const auto& tensor : list) {
+    current = prioritize(current, tensor, device_type);
+  }
+  return current;
+}
+
+// Template to catch non-Tensor args (no-op that returns current best guess)
+template <typename T>
+inline at::ScalarType prioritize(
+    at::ScalarType current,
+    T nextArg,
+    c10::DeviceType device_type = c10::DeviceType::CUDA) {
+  return current;
+}
+
+// Overload for the tail case.
+inline at::ScalarType promote_type(
+    at::ScalarType current,
+    c10::DeviceType device_type) {
+  return current;
+}
+
+// Unpack args and determine if incoming lower_precision_fp tensors need to be
+// promoted to float32. Non-Tensor arguments are ignored.
+template <typename Arg0, typename... Args>
+inline at::ScalarType promote_type(
+    at::ScalarType current,
+    c10::DeviceType device_type,
+    Arg0 arg0,
+    Args... args) {
+  auto new_current = prioritize(current, arg0, device_type);
+  return promote_type(new_current, device_type, args...);
+}
+
+/****************************************************
+Logic to apply cached casting to any Tensor argument.
+****************************************************/
+inline bool is_eligible(
+    const Tensor& arg,
+    c10::DeviceType device_type = c10::DeviceType::CUDA) {
+  return (
+      arg.defined() && is_autocast_eligible(arg, device_type) &&
+      (arg.scalar_type() != at::kDouble));
+}
+
+// Overload to catch Tensor args
+TORCH_API Tensor cached_cast(
+    at::ScalarType to_type,
+    const Tensor& arg,
+    c10::DeviceType device_type = c10::DeviceType::CUDA);
+
+// Overload to process optional<Tensor>
+inline c10::optional<Tensor> cached_cast(
+    at::ScalarType to_type,
+    const c10::optional<Tensor>& arg,
+    c10::DeviceType device_type = c10::DeviceType::CUDA) {
+  if (arg.has_value()) {
+    return cached_cast(to_type, *arg, device_type);
+  } else {
+    return c10::nullopt;
+  }
+}
+
+// Overload to process TensorLists
+inline std::vector<Tensor> cached_cast(
+    at::ScalarType to_type,
+    const TensorList& arg,
+    c10::DeviceType device_type = c10::DeviceType::CUDA) {
+  std::vector<Tensor> vec;
+  vec.reserve(arg.size());
+  for (const auto& t : arg) {
+    vec.emplace_back(cached_cast(to_type, t, device_type));
+  }
+  return vec;
+}
+
+inline std::vector<Tensor> cached_cast(
+    at::ScalarType to_type,
+    const ITensorListRef& arg,
+    c10::DeviceType device_type = c10::DeviceType::CUDA) {
+  std::vector<Tensor> vec;
+  vec.reserve(arg.size());
+  for (const auto& t : arg) {
+    vec.emplace_back(cached_cast(to_type, t, device_type));
+  }
+  return vec;
+}
+
+// Template to catch non-Tensor args.
+template <typename T>
+inline T cached_cast(
+    at::ScalarType to_type,
+    T arg,
+    c10::DeviceType device_type = c10::DeviceType::CUDA) {
+  return arg;
+}
+
+/*******************************************************
+Logic to flip an output dtype flag.
+Keep it simple for now by assuming only one such flag is
+present in the argument list.  If I ever need a function
+with more than flag I'll figure out something else.
+The policy is:
+If the user has explicity specified a dtype, respect it.
+Otherwise, set it to the autocast type.
+********************************************************/
+
+// Overload to catch dtype flags
+c10::optional<ScalarType> inline set_opt_dtype(
+    at::ScalarType to_type,
+    const c10::optional<ScalarType>& dtype) {
+  return dtype.has_value() ? dtype : to_type;
+}
+
+// Template to catch other args
+template <typename T>
+inline T set_opt_dtype(at::ScalarType to_type, T arg) {
+  return arg;
+}
+
+template <typename... Args>
+inline bool firstarg_is_eligible(
+    c10::DeviceType device_type,
+    const Tensor& arg,
+    Args... args) {
+  return is_eligible(arg, device_type);
+}
+
+template <typename... Args>
+inline at::ScalarType type_from_firstarg(
+    c10::DeviceType device_type,
+    at::ScalarType to_type,
+    const Tensor& arg,
+    Args... args) {
+  return (is_eligible(arg, device_type) ? to_type : arg.scalar_type());
+}
+
+// Policies correspond to op categories that need code-divergent handling.
+// Wrapper templates below are specialized based on a policy template parameter.
+enum class CastPolicy : uint8_t {
+  lower_precision_fp = 0, // Cast all inputs to lower_precision_fp before
+                          // running the op. Currently, lower_precision_fp is
+                          // fp16 for AutocastCUDA, and is defined by user
+                          // (default bf16) for AutocastCPU or other device.
+  fp32, // Cast all inputs to at::kFloat before running the op.
+  fp32_set_opt_dtype, // Treats functions (like softmax) that
+                      //  1. we'd like to run in fp32 and
+                      //  2. have a c10::optional<ScalarType> arg that controls
+                      //  the output type.
+                      // fp32_set_opt_dtype wrappers' policy is: if the output
+                      // type is already set, don't touch it, otherwise, set
+                      // it to at::kFloat.
+  fp32_append_dtype, // Treats functions (like norm) that
+                     //  1. we'd like to run in fp32 and
+                     //  2. have some overloads that accept an output type and
+                     //  other overloads that don't.
+                     // fp32_append_dtype wrappers wrap the overloads that don't
+                     // have an output dtype.
+                     // The wrapper policy is:  append at::kFloat to the args,
+                     // and redispatch to the type-aware overload.
+  promote, // Run in the widest dtype among several args.
+};
+
+/********************************************************************************************************
+Templates to provide wrapper functions
+
+I'm copying the pattern used in core/boxing/impl/WrapFunctionIntoFunctor.h to
+extract args and return type. (see also
+https://stackoverflow.com/questions/46533698/how-to-deduce-argument-list-from-function-pointer)
+
+This strategy uses an exterior "WrapFunction" that extracts arguments on behalf
+of (in my case several specializations of) an interior "WrapFunction_".
+Interior WrapFunction_ specializations are defined for each CastPolicy.
+********************************************************************************************************/
+
+// Base template for WrapFunction_, which is specialized to contain a "call"
+// method each CastPolicy
+template <
+    CastPolicy policy,
+    c10::DeviceType device_type,
+    class Redispatch,
+    Redispatch* F,
+    class Ret,
+    class ArgList>
+struct WrapFunction_ {};
+
+// CastPolicy::lower_precision_fp General_DeviceType
+template <
+    c10::DeviceType device_type,
+    class Redispatch,
+    Redispatch* F,
+    class Ret,
+    class... Args>
+struct WrapFunction_<
+    CastPolicy::lower_precision_fp,
+    device_type,
+    Redispatch,
+    F,
+    Ret,
+    guts::typelist::typelist<Args...>> {
+  static Ret call(Args... args) {
+    c10::impl::ExcludeDispatchKeyGuard no_autocast(
+        get_autocast_dispatch_key_from_device_type(device_type));
+    return (*F)(cached_cast(
+        get_lower_precision_fp_from_device_type(device_type),
+        args,
+        device_type)...);
+  }
+};
+
+// CastPolicy::fp32 General_DeviceType
+template <
+    c10::DeviceType device_type,
+    class Redispatch,
+    Redispatch* F,
+    class Ret,
+    class... Args>
+struct WrapFunction_<
+    CastPolicy::fp32,
+    device_type,
+    Redispatch,
+    F,
+    Ret,
+    guts::typelist::typelist<Args...>> {
+  static Ret call(Args... args) {
+    c10::impl::ExcludeDispatchKeyGuard no_autocast(
+        get_autocast_dispatch_key_from_device_type(device_type));
+    return (*F)(cached_cast(at::kFloat, args, device_type)...);
+  }
+};
+
+// CastPolicy::fp32_set_opt_dtype General_DeviceType
+template <
+    c10::DeviceType device_type,
+    class Redispatch,
+    Redispatch* F,
+    class Ret,
+    class... Args>
+struct WrapFunction_<
+    CastPolicy::fp32_set_opt_dtype,
+    device_type,
+    Redispatch,
+    F,
+    Ret,
+    guts::typelist::typelist<Args...>> {
+  static Ret call(Args... args) {
+    c10::impl::ExcludeDispatchKeyGuard no_autocast(
+        get_autocast_dispatch_key_from_device_type(device_type));
+    if (firstarg_is_eligible(device_type, args...)) {
+      return (*F)(set_opt_dtype(at::kFloat, args)...);
+    } else {
+      // If ineligible, calls F with unaltered args.  Does not set opt dtype,
+      // because setting opt dtype explicitly may interfere with internal
+      // implicit promotion decisions.
+      return (*F)(args...);
+    }
+  }
+};
+
+// CastPolicy::fp32_append_dtype General_DeviceType
+template <
+    c10::DeviceType device_type,
+    class Redispatch,
+    Redispatch* F,
+    class Ret,
+    class... Args>
+struct WrapFunction_<
+    CastPolicy::fp32_append_dtype,
+    device_type,
+    Redispatch,
+    F,
+    Ret,
+    guts::typelist::typelist<Args...>> {
+  static Ret call(Args... args) {
+    c10::impl::ExcludeDispatchKeyGuard no_autocast(
+        get_autocast_dispatch_key_from_device_type(device_type));
+    at::ScalarType out_type =
+        type_from_firstarg(device_type, at::kFloat, args...);
+    return (*F)(args..., out_type);
+  }
+};
+
+// CastPolicy::promote General_DeviceType
+template <
+    c10::DeviceType device_type,
+    class Redispatch,
+    Redispatch* F,
+    class Ret,
+    class... Args>
+struct WrapFunction_<
+    CastPolicy::promote,
+    device_type,
+    Redispatch,
+    F,
+    Ret,
+    guts::typelist::typelist<Args...>> {
+  static Ret call(Args... args) {
+    c10::impl::ExcludeDispatchKeyGuard no_autocast(
+        get_autocast_dispatch_key_from_device_type(device_type));
+    auto to_type = promote_type(
+        get_lower_precision_fp_from_device_type(device_type),
+        device_type,
+        args...);
+    return (*F)(cached_cast(to_type, args, device_type)...);
+  }
+};
+
+// Wrapper to infer return_type and parameter_types for WrapFunction_ (imitating
+// core/boxing/impl/WrapFunctionIntoFunctor.h)
+template <
+    CastPolicy policy,
+    c10::DeviceType device_type,
+    class Registered, // The signature for which we're registering.  The
+                      // dispatcher's calling code invokes our registered
+                      // functions with arguments matching Registered, so we
+                      // register WrapFunction_::call methods with a matching
+                      // signature to properly field those arguments.
+    // guts::function_traits below extracts return_type and
+    // parameter_types from Registered, which WrapFunction_
+    // templates above use to declare their call methods.
+    class Redispatch, // The signature for the function we're redispatching to.
+                      // In most cases this is the same as Registered, but for
+                      // some ops (for example, ops where we append a dtype)
+                      // it's useful to redispatch to a function with a
+                      // different signature.
+    Redispatch* F> // The actual function we're redispatching to.
+struct WrapFunction final {
+  using type = WrapFunction_<
+      policy,
+      device_type,
+      Redispatch,
+      F,
+      typename guts::function_traits<Registered>::return_type,
+      typename guts::function_traits<Registered>::parameter_types>;
+};
+
+/*****************************************************************************************************************
+This section performs load-time registration for autocast wrappers.
+
+It's debatable at what level operations should be patched.  We'd like casts to
+be autograd-exposed and precede autograd history recording, so that for
+lower_precision_fp ops, input tensors are saved for backward in
+lower_precision_fp rather than fp32.  Saving inputs in lower_precision_fp
+can significantly reduce a model's memory footprint.
+
+Option 1 (strawman):  Patch only at the level of explicit calls into
+cudnn/cublas (cudnn_convolution, etc), because those are the code paths that are
+guaranteed to use Tensor Cores, therefore they're the ones that will benefit
+most from lower_precision_fp.   Potential pitfall:  convolutions (and other ops)
+are wrapped in several layers of at::* calls.  If one of those happens to record
+autograd history, then we've lost the opportunity to save inputs in
+lower_precision_fp.
+
+Option 2:  Patch the Python-exposed surface of calls, to make 100% sure autograd
+history recording can't sneak in ahead of autocast.  This mirrors Apex most
+closely.
+
+I think Option 2 is the right answer for all ops, not just convolutions. Option
+2 is what I implement here.
+*****************************************************************************************************************/
+
+/********************************************************************************************************************
+Explicit registration for out-of-place ops
+
+The stuff below could be codegenned.  Ed said
+> you are going to have to write the function definition at some point, I
+wouldn't try to get clever about it Therefore, for the moment, this is all
+copy pasted in from VariableTypeEverything.cpp with appropriate substitutions.
+********************************************************************************************************************/
+
+} // namespace at::autocast
+
+#define ADD_NS(RAW_OP) at::RAW_OP
+
+// Common cases where registration signature matches redispatch signature
+// (that's why SIGNATURE is repeated in the WrapFunction instantiation)
+#define KERNEL(DISPATCHKEY, OP, POLICY)       \
+  m.impl(                                     \
+      TORCH_SELECTIVE_NAME("aten::" #OP),     \
+      &::at::autocast::WrapFunction<          \
+          ::at::autocast::CastPolicy::POLICY, \
+          DISPATCHKEY,                        \
+          decltype(ATEN_FN(OP)),              \
+          decltype(ATEN_FN(OP)),              \
+          &ATEN_FN(OP)>::type::call);
+
+#define KERNEL2(DISPATCHKEY, OP, OVERLOAD, POLICY)      \
+  m.impl(                                               \
+      TORCH_SELECTIVE_NAME("aten::" #OP "." #OVERLOAD), \
+      &::at::autocast::WrapFunction<                    \
+          ::at::autocast::CastPolicy::POLICY,           \
+          DISPATCHKEY,                                  \
+          decltype(ATEN_FN2(OP, OVERLOAD)),             \
+          decltype(ATEN_FN2(OP, OVERLOAD)),             \
+          &ATEN_FN2(OP, OVERLOAD)>::type::call);
+
+// Less-common but still useful case: redispatching to a function
+// with a new signature (e.g. appending a dtype)
+#define KERNEL_DIFFERENT_REDISPATCH_SIGNATURE(      \
+    DISPATCHKEY,                                    \
+    REDISPATCH_FUNC,                                \
+    REGISTER_NAME,                                  \
+    REGISTER_SIGNATURE,                             \
+    REDISPATCH_SIGNATURE,                           \
+    POLICY)                                         \
+  m.impl(                                           \
+      TORCH_SELECTIVE_NAME("aten::" REGISTER_NAME), \
+      &::at::autocast::WrapFunction<                \
+          ::at::autocast::CastPolicy::POLICY,       \
+          DISPATCHKEY,                              \
+          REGISTER_SIGNATURE,                       \
+          REDISPATCH_SIGNATURE,                     \
+          &REDISPATCH_FUNC>::type::call);
+
+// KERNEL_CPU/KERNEL_CPU2/KERNEL_DIFFERENT_REDISPATCH_SIGNATURE_CPU
+// registration for AutocastCPU
+#define KERNEL_CPU(OP, POLICY) KERNEL(c10::DeviceType::CPU, OP, POLICY)
+
+#define KERNEL_CPU2(OP, OVERLOAD, POLICY) \
+  KERNEL2(c10::DeviceType::CPU, OP, OVERLOAD, POLICY)
+
+#define KERNEL_DIFFERENT_REDISPATCH_SIGNATURE_CPU( \
+    REDISPATCH_FUNC,                               \
+    REGISTER_NAME,                                 \
+    REGISTER_SIGNATURE,                            \
+    REDISPATCH_SIGNATURE,                          \
+    POLICY)                                        \
+  KERNEL_DIFFERENT_REDISPATCH_SIGNATURE(           \
+      c10::DeviceType::CPU,                        \
+      REDISPATCH_FUNC,                             \
+      REGISTER_NAME,                               \
+      REGISTER_SIGNATURE,                          \
+      REDISPATCH_SIGNATURE,                        \
+      POLICY)
+
+// KERNEL_CUDA/KERNEL_CUDA2/KERNEL_DIFFERENT_REDISPATCH_SIGNATURE_CUDA
+// registration for AutocastCUDA
+#define KERNEL_CUDA(OP, POLICY) KERNEL(c10::DeviceType::CUDA, OP, POLICY)
+
+#define KERNEL_CUDA2(OP, OVERLOAD, POLICY) \
+  KERNEL2(c10::DeviceType::CUDA, OP, OVERLOAD, POLICY)
+
+#define KERNEL_DIFFERENT_REDISPATCH_SIGNATURE_CUDA( \
+    REDISPATCH_FUNC,                                \
+    REGISTER_NAME,                                  \
+    REGISTER_SIGNATURE,                             \
+    REDISPATCH_SIGNATURE,                           \
+    POLICY)                                         \
+  KERNEL_DIFFERENT_REDISPATCH_SIGNATURE(            \
+      c10::DeviceType::CUDA,                        \
+      REDISPATCH_FUNC,                              \
+      REGISTER_NAME,                                \
+      REGISTER_SIGNATURE,                           \
+      REDISPATCH_SIGNATURE,                         \
+      POLICY)
+
+// KERNEL_PRIVATEUSEONE/KERNEL_PRIVATEUSEONE2/
+// KERNEL_DIFFERENT_REDISPATCH_SIGNATURE_PRIVATEUSEONE
+// registration for AutocastPrivateUse1
+#define KERNEL_PRIVATEUSEONE(OP, POLICY) \
+  KERNEL(c10::DeviceType::PrivateUse1, OP, POLICY)
+
+#define KERNEL_PRIVATEUSEONE2(OP, OVERLOAD, POLICY) \
+  KERNEL2(c10::DeviceType::PrivateUse1, OP, OVERLOAD, POLICY)
+
+#define KERNEL_DIFFERENT_REDISPATCH_SIGNATURE_PRIVATEUSEONE( \
+    REDISPATCH_FUNC,                                         \
+    REGISTER_NAME,                                           \
+    REGISTER_SIGNATURE,                                      \
+    REDISPATCH_SIGNATURE,                                    \
+    POLICY)                                                  \
+  KERNEL_DIFFERENT_REDISPATCH_SIGNATURE(                     \
+      c10::DeviceType::PrivateUse1,                          \
+      REDISPATCH_FUNC,                                       \
+      REGISTER_NAME,                                         \
+      REGISTER_SIGNATURE,                                    \
+      REDISPATCH_SIGNATURE,                                  \
+      POLICY)

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/cpp_custom_type_hack.h

@@ -0,0 +1,110 @@
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+
+// YOU ARE IN THE WRONG PLACE! TURN BACK NOW!
+
+// This code was a temporary hack to enable embedding arbitrary C++ structures
+// into Tensors. THIS IS UNSAFE AND IS NOT SUPPORTED. IF YOU USE THIS CODE,
+// IT __WILL__ BREAK.
+
+// This code has been superseded by custom classes:
+// https://pytorch.org/tutorials/advanced/torch_script_custom_classes.html
+
+// Please use custom classes and **DO NOT ADD MORE CALLSITES TO THINGS DEFINED
+// IN THIS FILE**.
+
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+// STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP STOP
+
+#include <ATen/TracerMode.h>
+#include <ATen/core/Tensor.h>
+
+#ifndef AT_PER_OPERATOR_HEADERS
+#include <ATen/Functions.h>
+#else
+#include <ATen/ops/empty.h>
+#endif
+
+namespace at::cpp_custom_type_hack {
+
+template <typename T>
+[[deprecated(
+    "Use custom classes instead: "
+    "https://pytorch.org/tutorials/advanced/torch_script_custom_classes.html")]] bool
+isa(const Tensor& packed) {
+  return (packed.scalar_type() == kByte) &&
+      (packed.storage().data_ptr().get_deleter() ==
+       caffe2::TypeMeta::Make<T>().deleteFn());
+}
+
+template <typename T>
+[[deprecated(
+    "Use custom classes instead: "
+    "https://pytorch.org/tutorials/advanced/torch_script_custom_classes.html")]] T&
+cast(const Tensor& packed) {
+  TORCH_CHECK(
+      packed.scalar_type() == kByte, "Expected temporary cpp type wrapper");
+  TORCH_CHECK(
+      packed.storage().data_ptr().get_deleter() ==
+          caffe2::TypeMeta::Make<T>().deleteFn(),
+      "Expected temporary cpp type wrapper of type ",
+      caffe2::TypeMeta::TypeName<T>());
+  return *reinterpret_cast<T*>(packed.storage().data_ptr().get());
+}
+
+template <typename T>
+[[deprecated(
+    "Use custom classes instead: "
+    "https://pytorch.org/tutorials/advanced/torch_script_custom_classes.html")]] Tensor
+create(std::unique_ptr<T> ptr, TensorOptions options) {
+  // None of this should trace, so turn off Tracer dispatching
+  at::AutoDispatchBelowADInplaceOrView guard; // TODO: remove
+  at::tracer::impl::NoTracerDispatchMode tracer_guard;
+
+  // We store this instance away in a Tensor and register a deleter function
+  // so that we do not leak memory. On the other side, we pull out the storage's
+  // data_ptr and get the right typed pointer.
+  void* raw_ptr = ptr.release();
+  at::DataPtr at_ptr(
+      raw_ptr, raw_ptr, caffe2::TypeMeta::Make<T>().deleteFn(), at::kCPU);
+
+  // size doesn't really matter, but we can align it to the actual size
+  // returning variables because one likely want to use this hack from python
+  auto retval = at::empty({sizeof(T)}, options.device(kCPU).dtype(at::kByte));
+  retval.storage().set_data_ptr_noswap(std::move(at_ptr));
+  return retval;
+}
+
+} // namespace at::cpp_custom_type_hack

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/cuda/ApplyGridUtils.cuh

@@ -0,0 +1,47 @@
+#include <ATen/cuda/CUDAContext.h>
+
+#include <cuda_runtime.h>
+
+namespace at::cuda {
+
+/**
+   Computes ceil(a / b)
+*/
+template <typename T>
+__host__ __device__ __forceinline__ T ATenCeilDiv(T a, T b) {
+  return (a + b - 1) / b;
+}
+
+namespace {
+
+// Threads per block for our apply kernel
+// FIXME: use occupancy calculator instead
+constexpr uint32_t AT_APPLY_THREADS_PER_BLOCK = 512;
+constexpr uint32_t AT_APPLY_BLOCKS_PER_SM = 4;
+
+template <int step = 1>
+inline bool getApplyGrid(uint64_t totalElements, dim3& grid, c10::DeviceIndex curDevice, int max_threads_per_block=AT_APPLY_THREADS_PER_BLOCK) {
+  if (curDevice == -1) return false;
+  uint64_t numel_per_thread = static_cast<uint64_t>(max_threads_per_block) * static_cast<uint64_t>(step);
+  uint64_t numBlocks = ATenCeilDiv(totalElements, numel_per_thread);
+  uint64_t maxGridX = at::cuda::getDeviceProperties(curDevice)->maxGridSize[0];
+  if (numBlocks > maxGridX)
+    numBlocks = maxGridX;
+  grid = dim3(numBlocks);
+  return true;
+}
+
+constexpr int getApplyBlocksPerSM() {
+  return AT_APPLY_BLOCKS_PER_SM;
+}
+
+constexpr int getApplyBlockSize() {
+  return AT_APPLY_THREADS_PER_BLOCK;
+}
+
+inline dim3 getApplyBlock(int max_threads_per_block=AT_APPLY_THREADS_PER_BLOCK) {
+  return dim3(max_threads_per_block);
+}
+
+} // anonymous namespace
+} // namespace at::cuda

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/cuda/AsmUtils.cuh

@@ -0,0 +1,149 @@
+#pragma once
+#include <cstdint>
+
+// Collection of direct PTX functions
+
+namespace at::cuda {
+
+template <typename T>
+struct Bitfield {};
+
+template <>
+struct Bitfield<unsigned int> {
+  static __device__ __host__ __forceinline__
+  unsigned int getBitfield(unsigned int val, int pos, int len) {
+#if !defined(__CUDA_ARCH__)
+    pos &= 0xff;
+    len &= 0xff;
+
+    unsigned int m = (1u << len) - 1u;
+    return (val >> pos) & m;
+#else
+    unsigned int ret;
+    asm("bfe.u32 %0, %1, %2, %3;" : "=r"(ret) : "r"(val), "r"(pos), "r"(len));
+    return ret;
+#endif
+  }
+
+  static __device__ __host__ __forceinline__
+  unsigned int setBitfield(unsigned int val, unsigned int toInsert, int pos, int len) {
+#if !defined(__CUDA_ARCH__)
+    pos &= 0xff;
+    len &= 0xff;
+
+    unsigned int m = (1u << len) - 1u;
+    toInsert &= m;
+    toInsert <<= pos;
+    m <<= pos;
+
+    return (val & ~m) | toInsert;
+#else
+    unsigned int ret;
+    asm("bfi.b32 %0, %1, %2, %3, %4;" :
+        "=r"(ret) : "r"(toInsert), "r"(val), "r"(pos), "r"(len));
+    return ret;
+#endif
+  }
+};
+
+template <>
+struct Bitfield<uint64_t> {
+  static __device__ __host__ __forceinline__
+  uint64_t getBitfield(uint64_t val, int pos, int len) {
+#if !defined(__CUDA_ARCH__)
+    pos &= 0xff;
+    len &= 0xff;
+
+    uint64_t m = (1u << len) - 1u;
+    return (val >> pos) & m;
+#else
+    uint64_t ret;
+    asm("bfe.u64 %0, %1, %2, %3;" : "=l"(ret) : "l"(val), "r"(pos), "r"(len));
+    return ret;
+#endif
+  }
+
+  static __device__ __host__ __forceinline__
+  uint64_t setBitfield(uint64_t val, uint64_t toInsert, int pos, int len) {
+#if !defined(__CUDA_ARCH__)
+    pos &= 0xff;
+    len &= 0xff;
+
+    uint64_t m = (1u << len) - 1u;
+    toInsert &= m;
+    toInsert <<= pos;
+    m <<= pos;
+
+    return (val & ~m) | toInsert;
+#else
+    uint64_t ret;
+    asm("bfi.b64 %0, %1, %2, %3, %4;" :
+        "=l"(ret) : "l"(toInsert), "l"(val), "r"(pos), "r"(len));
+    return ret;
+#endif
+  }
+};
+
+__device__ __forceinline__ int getLaneId() {
+#if defined(USE_ROCM)
+  return __lane_id();
+#else
+  int laneId;
+  asm("mov.s32 %0, %%laneid;" : "=r"(laneId) );
+  return laneId;
+#endif
+}
+
+#if defined(USE_ROCM)
+__device__ __forceinline__ unsigned long long int getLaneMaskLt() {
+  const std::uint64_t m = (1ull << getLaneId()) - 1ull;
+  return m;
+}
+#else
+__device__ __forceinline__ unsigned getLaneMaskLt() {
+  unsigned mask;
+  asm("mov.u32 %0, %%lanemask_lt;" : "=r"(mask));
+  return mask;
+}
+#endif
+
+#if defined (USE_ROCM)
+__device__ __forceinline__ unsigned long long int getLaneMaskLe() {
+  std::uint64_t m = UINT64_MAX >> (sizeof(std::uint64_t) * CHAR_BIT - (getLaneId() + 1));
+  return m;
+}
+#else
+__device__ __forceinline__ unsigned getLaneMaskLe() {
+  unsigned mask;
+  asm("mov.u32 %0, %%lanemask_le;" : "=r"(mask));
+  return mask;
+}
+#endif
+
+#if defined(USE_ROCM)
+__device__ __forceinline__ unsigned long long int getLaneMaskGt() {
+  const std::uint64_t m = getLaneMaskLe();
+  return m ? ~m : m;
+}
+#else
+__device__ __forceinline__ unsigned getLaneMaskGt() {
+  unsigned mask;
+  asm("mov.u32 %0, %%lanemask_gt;" : "=r"(mask));
+  return mask;
+}
+#endif
+
+#if defined(USE_ROCM)
+__device__ __forceinline__ unsigned long long int getLaneMaskGe() {
+  const std::uint64_t m = getLaneMaskLt();
+  return ~m;
+}
+#else
+__device__ __forceinline__ unsigned getLaneMaskGe() {
+  unsigned mask;
+  asm("mov.u32 %0, %%lanemask_ge;" : "=r"(mask));
+  return mask;
+}
+#endif
+
+} // namespace at::cuda

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/cuda/CUDABlas.h

@@ -0,0 +1,375 @@
+#pragma once
+/*
+  Provides a subset of CUDA BLAS functions as templates:
+
+    gemm<Dtype>(transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c,
+  ldc)
+
+    gemv<Dtype>(transa, m, n, alpha, a, lda, x, incx, beta, y, incy)
+
+    dot<Dtype>(n, x, incx, y, incy, result)
+
+  where Dtype is double, float, at::Half or at::BFloat16 (ROCm, NOT for dot).
+  The functions are available in at::cuda::blas namespace.
+ */
+
+#include <ATen/cuda/CUDAContext.h>
+#include <ATen/OpMathType.h>
+
+namespace at::cuda::blas {
+
+// RAII guard that sets the CuBLAS pointer mode and restores it to
+// its previous value when the guard is destroyed
+class PointerModeGuard {
+public:
+  PointerModeGuard(cublasHandle_t handle, cublasPointerMode_t mode) :
+      handle(handle) {
+    TORCH_CUDABLAS_CHECK(cublasGetPointerMode(handle, &previous_mode));
+    TORCH_CUDABLAS_CHECK(cublasSetPointerMode(handle, mode));
+  }
+
+  ~PointerModeGuard() {
+    cublasSetPointerMode(handle, previous_mode);
+  }
+
+private:
+  cublasHandle_t handle;
+  cublasPointerMode_t previous_mode;
+};
+
+/* LEVEL 3 BLAS FUNCTIONS */
+
+#define CUDABLAS_GEMM_ARGTYPES(Dtype)                                                       \
+  char transa, char transb, int64_t m, int64_t n, int64_t k, at::opmath_type<Dtype> alpha,  \
+      const Dtype *a, int64_t lda, const Dtype *b, int64_t ldb, at::opmath_type<Dtype> beta,\
+      Dtype *c, int64_t ldc
+
+#define CUDABLAS_GEMM_ARGS(Dtype) transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc
+
+template <typename Dtype>
+inline void gemm(CUDABLAS_GEMM_ARGTYPES(Dtype)) {
+  AT_ERROR("at::cuda::blas::gemm: not implemented for ", typeid(Dtype).name());
+}
+
+template <>
+void gemm<double>(CUDABLAS_GEMM_ARGTYPES(double));
+template <>
+void gemm<float>(CUDABLAS_GEMM_ARGTYPES(float));
+template <>
+void gemm<c10::complex<double>>(CUDABLAS_GEMM_ARGTYPES(c10::complex<double>));
+template <>
+void gemm<c10::complex<float>>(CUDABLAS_GEMM_ARGTYPES(c10::complex<float>));
+template <>
+void gemm<at::Half>(CUDABLAS_GEMM_ARGTYPES(at::Half));
+template <>
+void gemm<at::BFloat16>(CUDABLAS_GEMM_ARGTYPES(at::BFloat16));
+
+template <typename Dtype>
+inline void gemm_internal(CUDABLAS_GEMM_ARGTYPES(Dtype)) {
+  AT_ERROR("at::cuda::blas::gemm_internal: not implemented for ", typeid(Dtype).name());
+}
+
+template <>
+void gemm_internal<double>(CUDABLAS_GEMM_ARGTYPES(double));
+template <>
+void gemm_internal<float>(CUDABLAS_GEMM_ARGTYPES(float));
+template <>
+void gemm_internal<c10::complex<double>>(CUDABLAS_GEMM_ARGTYPES(c10::complex<double>));
+template <>
+void gemm_internal<c10::complex<float>>(CUDABLAS_GEMM_ARGTYPES(c10::complex<float>));
+template <>
+void gemm_internal<at::Half>(CUDABLAS_GEMM_ARGTYPES(at::Half));
+template <>
+void gemm_internal<at::BFloat16>(CUDABLAS_GEMM_ARGTYPES(at::BFloat16));
+
+#if (!defined(USE_ROCM) && !defined(_MSC_VER)) || (defined(USE_ROCM) && ROCM_VERSION >= 50700)
+enum GEMMAndBiasActivationEpilogue {
+  None,
+  RELU,
+  GELU,
+};
+
+// NOTE: GELU activation is not supported prior to CUDA 11.4 and will
+// do nothing if passed in that case.
+template <typename Dtype>
+void gemm_and_bias(
+    bool transpose_mat1,
+    bool transpose_mat2,
+    int64_t m,
+    int64_t n,
+    int64_t k,
+    at::opmath_type<Dtype> alpha_val,
+    const Dtype* mat1_ptr,
+    int64_t mat1_ld,
+    const Dtype* mat2_ptr,
+    int64_t mat2_ld,
+    const Dtype* bias,
+    Dtype* result_ptr,
+    int64_t result_ld,
+    GEMMAndBiasActivationEpilogue activation = GEMMAndBiasActivationEpilogue::None);
+
+void int8_gemm(
+    bool transpose_mat1,
+    bool transpose_mat2,
+    int64_t m,
+    int64_t n,
+    int64_t k,
+    const int8_t* mat1_ptr,
+    int64_t mat1_ld,
+    const int8_t* mat2_ptr,
+    int64_t mat2_ld,
+    int32_t* result_ptr,
+    int64_t result_ld);
+
+void scaled_gemm(
+    char transa,
+    char transb,
+    int64_t m,
+    int64_t n,
+    int64_t k,
+    const void* mat1_ptr,
+    const void* mat1_scale_ptr,
+    int64_t mat1_ld,
+    ScalarType mat1_dtype,
+    const void* mat2_ptr,
+    const void* mat2_scale_ptr,
+    int64_t mat2_ld,
+    ScalarType mat2_dtype,
+    const void* bias_ptr,
+    ScalarType bias_dtype,
+    void* result_ptr,
+    const void* result_scale_ptr,
+    int64_t result_ld,
+    ScalarType result_dtype,
+    void* amax_ptr,
+    bool use_fast_accum);
+#endif
+
+#define CUDABLAS_BGEMM_ARGTYPES(Dtype)                                                        \
+  char transa, char transb, int64_t m, int64_t n, int64_t k, at::opmath_type<Dtype> alpha,    \
+      const Dtype *a, int64_t lda, int64_t stridea,                                           \
+      const Dtype *b, int64_t ldb, int64_t strideb,                                           \
+      at::opmath_type<Dtype> beta, Dtype *c, int64_t ldc, int64_t stridec, int64_t num_batches
+
+#define CUDABLAS_BGEMM_ARGS(Dtype) \
+  transa, transb, m, n, k, alpha, a, lda, stridea, b, ldb, strideb, beta, c, ldc, stridec, num_batches
+
+template <typename Dtype>
+inline void bgemm(CUDABLAS_BGEMM_ARGTYPES(Dtype)) {
+  AT_ERROR("at::cuda::blas::bgemm: not implemented for ", typeid(Dtype).name());
+}
+
+template <>
+void bgemm<double>(CUDABLAS_BGEMM_ARGTYPES(double));
+template <>
+void bgemm<float>(CUDABLAS_BGEMM_ARGTYPES(float));
+template <>
+void bgemm<c10::complex<double>>(CUDABLAS_BGEMM_ARGTYPES(c10::complex<double>));
+template <>
+void bgemm<c10::complex<float>>(CUDABLAS_BGEMM_ARGTYPES(c10::complex<float>));
+template <>
+void bgemm<at::Half>(CUDABLAS_BGEMM_ARGTYPES(at::Half));
+template <>
+void bgemm<at::BFloat16>(CUDABLAS_BGEMM_ARGTYPES(at::BFloat16));
+
+template <typename Dtype>
+inline void bgemm_internal(CUDABLAS_BGEMM_ARGTYPES(Dtype)) {
+  AT_ERROR("at::cuda::blas::bgemm_internal: not implemented for ", typeid(Dtype).name());
+}
+
+template <>
+void bgemm_internal<double>(CUDABLAS_BGEMM_ARGTYPES(double));
+template <>
+void bgemm_internal<float>(CUDABLAS_BGEMM_ARGTYPES(float));
+template <>
+void bgemm_internal<c10::complex<double>>(CUDABLAS_BGEMM_ARGTYPES(c10::complex<double>));
+template <>
+void bgemm_internal<c10::complex<float>>(CUDABLAS_BGEMM_ARGTYPES(c10::complex<float>));
+template <>
+void bgemm_internal<at::Half>(CUDABLAS_BGEMM_ARGTYPES(at::Half));
+template <>
+void bgemm_internal<at::BFloat16>(CUDABLAS_BGEMM_ARGTYPES(at::BFloat16));
+
+#if defined(USE_ROCM) && ROCM_VERSION <= 50500
+// ROCm 5.6 hipblas matches the const Dtype *A API, but prior hipblas does not.
+#define CUDABLAS_TRSM_ARGTYPES(Dtype)                                  \
+  hipblasHandle_t handle, hipblasSideMode_t side, hipblasFillMode_t uplo, \
+      hipblasOperation_t trans, hipblasDiagType_t diag, int m, int n,    \
+      const Dtype *alpha,       Dtype *A, int lda, Dtype *B, int ldb
+#else
+#define CUDABLAS_TRSM_ARGTYPES(Dtype)                                  \
+  cublasHandle_t handle, cublasSideMode_t side, cublasFillMode_t uplo, \
+      cublasOperation_t trans, cublasDiagType_t diag, int m, int n,    \
+      const Dtype *alpha, const Dtype *A, int lda, Dtype *B, int ldb
+#endif
+
+template <typename Dtype>
+inline void trsm(CUDABLAS_TRSM_ARGTYPES(Dtype)) {
+  TORCH_INTERNAL_ASSERT(false, "at::cuda::blas::trsm: not implemented for ", typeid(Dtype).name());
+}
+
+template <>
+TORCH_CUDA_CU_API void trsm<float>(CUDABLAS_TRSM_ARGTYPES(float));
+template <>
+TORCH_CUDA_CU_API void trsm<double>(CUDABLAS_TRSM_ARGTYPES(double));
+template <>
+TORCH_CUDA_CU_API void trsm<c10::complex<float>>(CUDABLAS_TRSM_ARGTYPES(c10::complex<float>));
+template <>
+TORCH_CUDA_CU_API void trsm<c10::complex<double>>(CUDABLAS_TRSM_ARGTYPES(c10::complex<double>));
+
+#define CUDABLAS_TRSM_BATCHED_ARGTYPES(Dtype)                          \
+  cublasHandle_t handle, cublasSideMode_t side, cublasFillMode_t uplo, \
+      cublasOperation_t trans, cublasDiagType_t diag, int m, int n,    \
+      const Dtype *alpha, Dtype *A[], int lda, Dtype *B[], int ldb,    \
+      int batchCount
+
+template <typename Dtype>
+inline void trsmBatched(CUDABLAS_TRSM_BATCHED_ARGTYPES(Dtype)) {
+  TORCH_INTERNAL_ASSERT(
+      false,
+      "at::cuda::blas::trsmBatched: not implemented for ",
+      typeid(Dtype).name());
+}
+
+template <>
+TORCH_CUDA_CU_API void trsmBatched<float>(CUDABLAS_TRSM_BATCHED_ARGTYPES(float));
+template <>
+TORCH_CUDA_CU_API void trsmBatched<double>(CUDABLAS_TRSM_BATCHED_ARGTYPES(double));
+template <>
+TORCH_CUDA_CU_API void trsmBatched<c10::complex<float>>(CUDABLAS_TRSM_BATCHED_ARGTYPES(c10::complex<float>));
+template <>
+TORCH_CUDA_CU_API void trsmBatched<c10::complex<double>>(CUDABLAS_TRSM_BATCHED_ARGTYPES(c10::complex<double>));
+
+/* LEVEL 2 BLAS FUNCTIONS */
+
+#define CUDABLAS_GEMV_ARGTYPES(Dtype)                                         \
+  char trans, int64_t m, int64_t n, Dtype alpha, const Dtype *a, int64_t lda, \
+      const Dtype *x, int64_t incx, Dtype beta, Dtype *y, int64_t incy
+
+template <typename Dtype>
+inline void gemv(CUDABLAS_GEMV_ARGTYPES(Dtype)) {
+  AT_ERROR("at::cuda::blas::gemv: not implemented for ", typeid(Dtype).name());
+}
+
+template <>
+void gemv<double>(CUDABLAS_GEMV_ARGTYPES(double));
+template <>
+void gemv<float>(CUDABLAS_GEMV_ARGTYPES(float));
+template <>
+void gemv<c10::complex<double>>(CUDABLAS_GEMV_ARGTYPES(c10::complex<double>));
+template <>
+void gemv<c10::complex<float>>(CUDABLAS_GEMV_ARGTYPES(c10::complex<float>));
+template <>
+void gemv<at::Half>(CUDABLAS_GEMV_ARGTYPES(at::Half));
+template <>
+void gemv<at::BFloat16>(CUDABLAS_GEMV_ARGTYPES(at::BFloat16));
+
+/* LEVEL 1 BLAS FUNCTIONS */
+
+#define CUDABLAS_DOT_ARGTYPES(Dtype)                                      \
+  cublasHandle_t handle, int n, const Dtype *x, int incx, const Dtype *y, \
+      int incy, Dtype *result
+
+template <typename Dtype>
+inline void dot(CUDABLAS_DOT_ARGTYPES(Dtype)) {
+  AT_ERROR("at::cuda::blas::dot: not implemented for ", typeid(Dtype).name());
+}
+
+template <>
+void dot<double>(CUDABLAS_DOT_ARGTYPES(double));
+template <>
+void dot<float>(CUDABLAS_DOT_ARGTYPES(float));
+template <>
+void dot<at::Half>(CUDABLAS_DOT_ARGTYPES(at::Half));
+template <>
+void dot<at::BFloat16>(CUDABLAS_DOT_ARGTYPES(at::BFloat16));
+template <>
+void dot<c10::complex<double>>(CUDABLAS_DOT_ARGTYPES(c10::complex<double>));
+template <>
+void dot<c10::complex<float>>(CUDABLAS_DOT_ARGTYPES(c10::complex<float>));
+
+template <typename Dtype>
+inline void vdot(CUDABLAS_DOT_ARGTYPES(Dtype)) {
+  AT_ERROR("at::cuda::blas::vdot: not implemented for ", typeid(Dtype).name());
+}
+
+template <>
+void vdot<c10::complex<float>>(CUDABLAS_DOT_ARGTYPES(c10::complex<float>));
+template <>
+void vdot<c10::complex<double>>(CUDABLAS_DOT_ARGTYPES(c10::complex<double>));
+
+#define CUDABLAS_GETRS_ARGTYPES(Dtype)  \
+  cublasHandle_t handle, cublasOperation_t trans, \
+  int n, int nrhs, Dtype** dA_array, int lda, int* ipiv_array, \
+  Dtype** dB_array, int ldb, int* info_array, int batchsize
+
+template<class Dtype>
+void getrsBatched(CUDABLAS_GETRS_ARGTYPES(Dtype)) {
+  TORCH_INTERNAL_ASSERT(false, "at::cuda::blas::getrsBatched: not implemented for ",
+    typeid(Dtype).name());
+}
+template<>
+TORCH_CUDA_CU_API void getrsBatched<float>(CUDABLAS_GETRS_ARGTYPES(float));
+template<>
+TORCH_CUDA_CU_API void getrsBatched<double>(CUDABLAS_GETRS_ARGTYPES(double));
+template<>
+TORCH_CUDA_CU_API void getrsBatched<c10::complex<float>>(CUDABLAS_GETRS_ARGTYPES(c10::complex<float>));
+template<>
+TORCH_CUDA_CU_API void getrsBatched<c10::complex<double>>(CUDABLAS_GETRS_ARGTYPES(c10::complex<double>));
+
+#define CUDABLAS_GEQRF_BATCHED_ARGTYPES(Dtype)                   \
+  cublasHandle_t handle, int m, int n, Dtype **A_array, int lda, \
+      Dtype **tau_array, int *info, int batchsize
+
+template <class Dtype>
+void geqrfBatched(CUDABLAS_GEQRF_BATCHED_ARGTYPES(Dtype)) {
+  TORCH_INTERNAL_ASSERT(
+      false,
+      "at::cuda::blas::geqrfBatched: not implemented for ",
+      typeid(Dtype).name());
+}
+template <>
+TORCH_CUDA_CU_API void geqrfBatched<float>(CUDABLAS_GEQRF_BATCHED_ARGTYPES(float));
+template <>
+TORCH_CUDA_CU_API void geqrfBatched<double>(CUDABLAS_GEQRF_BATCHED_ARGTYPES(double));
+template <>
+TORCH_CUDA_CU_API void geqrfBatched<c10::complex<double>>(
+    CUDABLAS_GEQRF_BATCHED_ARGTYPES(c10::complex<double>));
+template <>
+TORCH_CUDA_CU_API void geqrfBatched<c10::complex<float>>(
+    CUDABLAS_GEQRF_BATCHED_ARGTYPES(c10::complex<float>));
+
+#define CUDABLAS_GETRF_ARGTYPES(Dtype)  \
+  int n, Dtype** dA_array, int ldda, int* ipiv_array, int* info_array, int batchsize
+
+template<class Dtype>
+void getrfBatched(CUDABLAS_GETRF_ARGTYPES(Dtype)) {
+  TORCH_CHECK(false, "at::cuda::blas::getrfBatched: not implemented for ", typeid(Dtype).name());
+}
+template<>
+TORCH_CUDA_CU_API void getrfBatched<float>(CUDABLAS_GETRF_ARGTYPES(float));
+template<>
+TORCH_CUDA_CU_API void getrfBatched<double>(CUDABLAS_GETRF_ARGTYPES(double));
+template<>
+TORCH_CUDA_CU_API void getrfBatched<c10::complex<double>>(CUDABLAS_GETRF_ARGTYPES(c10::complex<double>));
+template<>
+TORCH_CUDA_CU_API void getrfBatched<c10::complex<float>>(CUDABLAS_GETRF_ARGTYPES(c10::complex<float>));
+
+#define CUDABLAS_GELS_BATCHED_ARGTYPES(Dtype)  \
+  cublasHandle_t handle, cublasOperation_t trans, int m, int n, int nrhs, Dtype** dA_array, int ldda, Dtype** dC_array, int lddc, int* info, int *devInfoArray, int batchSize
+
+template <class Dtype>
+void gelsBatched(CUDABLAS_GELS_BATCHED_ARGTYPES(Dtype)) {
+  TORCH_INTERNAL_ASSERT(false, "at::cuda::blas::gelsBatched: not implemented for ", typeid(Dtype).name());
+}
+
+template<>
+TORCH_CUDA_CU_API void gelsBatched<double>(CUDABLAS_GELS_BATCHED_ARGTYPES(double));
+template<>
+TORCH_CUDA_CU_API void gelsBatched<float>(CUDABLAS_GELS_BATCHED_ARGTYPES(float));
+template<>
+TORCH_CUDA_CU_API void gelsBatched<c10::complex<double>>(CUDABLAS_GELS_BATCHED_ARGTYPES(c10::complex<double>));
+template<>
+TORCH_CUDA_CU_API void gelsBatched<c10::complex<float>>(CUDABLAS_GELS_BATCHED_ARGTYPES(c10::complex<float>));
+
+} // namespace at::cuda::blas

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/cuda/CUDAConfig.h

@@ -0,0 +1,19 @@
+#pragma once
+
+// Test these using #if AT_CUDNN_ENABLED(), not #ifdef, so that it's
+// obvious if you forgot to include Config.h
+//    c.f. https://stackoverflow.com/questions/33759787/generating-an-error-if-checked-boolean-macro-is-not-defined
+//
+// NB: This header MUST NOT be included from other headers; it should
+// only be included from C++ files.
+#define AT_CUDNN_ENABLED() 1
+#define AT_CUSPARSELT_ENABLED() 1
+#define AT_ROCM_ENABLED() 0
+#define AT_MAGMA_ENABLED() 1
+
+// Needed for hipMAGMA to correctly identify implementation
+#if (AT_ROCM_ENABLED() && AT_MAGMA_ENABLED())
+#define HAVE_HIP 1
+#endif
+
+#define NVCC_FLAGS_EXTRA "-gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_75,code=sm_75;-gencode;arch=compute_80,code=sm_80;-gencode;arch=compute_86,code=sm_86;-gencode;arch=compute_37,code=sm_37;-gencode;arch=compute_90,code=sm_90"

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/cuda/CUDAContextLight.h

@@ -0,0 +1,95 @@
+#pragma once
+// Light-weight version of CUDAContext.h with fewer transitive includes
+
+#include <cstdint>
+
+#include <cuda_runtime_api.h>
+#include <cusparse.h>
+#include <cublas_v2.h>
+
+// cublasLT was introduced in CUDA 10.1 but we enable only for 11.1 that also
+// added bf16 support
+#if (!defined(USE_ROCM) && !defined(_MSC_VER)) || (defined(USE_ROCM) && ROCM_VERSION >= 50700)
+#include <cublasLt.h>
+#endif
+
+#ifdef CUDART_VERSION
+#include <cusolverDn.h>
+#endif
+
+#if defined(USE_ROCM) && ROCM_VERSION >= 50300
+#include <hipsolver/hipsolver.h>
+#endif
+
+#include <c10/core/Allocator.h>
+#include <c10/cuda/CUDAFunctions.h>
+
+namespace c10 {
+struct Allocator;
+}
+
+namespace at::cuda {
+
+/*
+A common CUDA interface for ATen.
+
+This interface is distinct from CUDAHooks, which defines an interface that links
+to both CPU-only and CUDA builds. That interface is intended for runtime
+dispatch and should be used from files that are included in both CPU-only and
+CUDA builds.
+
+CUDAContext, on the other hand, should be preferred by files only included in
+CUDA builds. It is intended to expose CUDA functionality in a consistent
+manner.
+
+This means there is some overlap between the CUDAContext and CUDAHooks, but
+the choice of which to use is simple: use CUDAContext when in a CUDA-only file,
+use CUDAHooks otherwise.
+
+Note that CUDAContext simply defines an interface with no associated class.
+It is expected that the modules whose functions compose this interface will
+manage their own state. There is only a single CUDA context/state.
+*/
+
+/**
+ * DEPRECATED: use device_count() instead
+ */
+inline int64_t getNumGPUs() {
+    return c10::cuda::device_count();
+}
+
+/**
+ * CUDA is available if we compiled with CUDA, and there are one or more
+ * devices.  If we compiled with CUDA but there is a driver problem, etc.,
+ * this function will report CUDA is not available (rather than raise an error.)
+ */
+inline bool is_available() {
+    return c10::cuda::device_count() > 0;
+}
+
+TORCH_CUDA_CPP_API cudaDeviceProp* getCurrentDeviceProperties();
+
+TORCH_CUDA_CPP_API int warp_size();
+
+TORCH_CUDA_CPP_API cudaDeviceProp* getDeviceProperties(c10::DeviceIndex device);
+
+TORCH_CUDA_CPP_API bool canDeviceAccessPeer(
+    c10::DeviceIndex device,
+    c10::DeviceIndex peer_device);
+
+TORCH_CUDA_CPP_API c10::Allocator* getCUDADeviceAllocator();
+
+/* Handles */
+TORCH_CUDA_CPP_API cusparseHandle_t getCurrentCUDASparseHandle();
+TORCH_CUDA_CPP_API cublasHandle_t getCurrentCUDABlasHandle();
+#if (!defined(USE_ROCM) && !defined(_MSC_VER)) || (defined(USE_ROCM) && ROCM_VERSION >= 50700)
+TORCH_CUDA_CPP_API cublasLtHandle_t getCurrentCUDABlasLtHandle();
+#endif
+
+TORCH_CUDA_CPP_API void clearCublasWorkspaces();
+
+#if defined(CUDART_VERSION) || defined(USE_ROCM) && ROCM_VERSION >= 50300
+TORCH_CUDA_CPP_API cusolverDnHandle_t getCurrentCUDASolverDnHandle();
+#endif
+
+} // namespace at::cuda

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/cuda/CUDADevice.h

@@ -0,0 +1,23 @@
+#pragma once
+
+#include <ATen/cuda/Exceptions.h>
+
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+namespace at::cuda {
+
+inline Device getDeviceFromPtr(void* ptr) {
+  cudaPointerAttributes attr{};
+
+  AT_CUDA_CHECK(cudaPointerGetAttributes(&attr, ptr));
+
+#if !defined(USE_ROCM)
+  TORCH_CHECK(attr.type != cudaMemoryTypeUnregistered,
+    "The specified pointer resides on host memory and is not registered with any CUDA device.");
+#endif
+
+  return {c10::DeviceType::CUDA, static_cast<DeviceIndex>(attr.device)};
+}
+
+} // namespace at::cuda

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/cuda/CUDAEvent.h

@@ -0,0 +1,208 @@
+#pragma once
+
+#include <ATen/cuda/ATenCUDAGeneral.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/core/impl/GPUTrace.h>
+#include <c10/cuda/CUDAStream.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <ATen/cuda/Exceptions.h>
+#include <c10/util/Exception.h>
+
+#include <cuda_runtime_api.h>
+
+#include <cstdint>
+#include <utility>
+
+namespace at::cuda {
+
+/*
+* CUDAEvents are movable not copyable wrappers around CUDA's events.
+*
+* CUDAEvents are constructed lazily when first recorded unless it is
+* reconstructed from a cudaIpcEventHandle_t. The event has a device, and this
+* device is acquired from the first recording stream. However, if reconstructed
+* from a handle, the device should be explicitly specified; or if ipc_handle() is
+* called before the event is ever recorded, it will use the current device.
+* Later streams that record the event must match this device.
+*/
+struct TORCH_CUDA_CPP_API CUDAEvent {
+  // Constructors
+  // Default value for `flags` is specified below - it's cudaEventDisableTiming
+  CUDAEvent() noexcept = default;
+  CUDAEvent(unsigned int flags) noexcept : flags_{flags} {}
+
+  CUDAEvent(
+      DeviceIndex device_index, const cudaIpcEventHandle_t* handle) {
+      device_index_ = device_index;
+      CUDAGuard guard(device_index_);
+
+      AT_CUDA_CHECK(cudaIpcOpenEventHandle(&event_, *handle));
+      is_created_ = true;
+  }
+
+  // Note: event destruction done on creating device to avoid creating a
+  // CUDA context on other devices.
+  ~CUDAEvent() {
+    try {
+      if (is_created_) {
+        CUDAGuard guard(device_index_);
+        const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
+        if (C10_UNLIKELY(interp)) {
+          (*interp)->trace_gpu_event_deletion(reinterpret_cast<uintptr_t>(event_));
+        }
+        AT_CUDA_CHECK(cudaEventDestroy(event_));
+      }
+    } catch (...) { /* No throw */ }
+  }
+
+  CUDAEvent(const CUDAEvent&) = delete;
+  CUDAEvent& operator=(const CUDAEvent&) = delete;
+
+  CUDAEvent(CUDAEvent&& other) noexcept { moveHelper(std::move(other)); }
+  CUDAEvent& operator=(CUDAEvent&& other) noexcept {
+    if (this != &other) {
+      moveHelper(std::move(other));
+    }
+    return *this;
+  }
+
+  operator cudaEvent_t() const { return event(); }
+
+  // Less than operator (to allow use in sets)
+  friend bool operator<(const CUDAEvent& left, const CUDAEvent& right) {
+    return left.event_ < right.event_;
+  }
+
+  optional<at::Device> device() const {
+    if (is_created_) {
+      return at::Device(at::kCUDA, device_index_);
+    } else {
+      return {};
+    }
+  }
+
+  bool isCreated() const { return is_created_; }
+  DeviceIndex device_index() const {return device_index_;}
+  cudaEvent_t event() const { return event_; }
+
+  // Note: cudaEventQuery can be safely called from any device
+  bool query() const {
+    if (!is_created_) {
+      return true;
+    }
+
+    cudaError_t err = cudaEventQuery(event_);
+    if (err == cudaSuccess) {
+      return true;
+    } else if (err != cudaErrorNotReady) {
+      C10_CUDA_CHECK(err);
+    } else {
+      // ignore and clear the error if not ready
+      (void)cudaGetLastError();
+    }
+
+    return false;
+  }
+
+  void record() { record(getCurrentCUDAStream()); }
+
+  void recordOnce(const CUDAStream& stream) {
+    if (!was_recorded_) record(stream);
+  }
+
+  // Note: cudaEventRecord must be called on the same device as the event.
+  void record(const CUDAStream& stream) {
+    if (!is_created_) {
+      createEvent(stream.device_index());
+    }
+
+    TORCH_CHECK(device_index_ == stream.device_index(), "Event device ", device_index_,
+      " does not match recording stream's device ", stream.device_index(), ".");
+    CUDAGuard guard(device_index_);
+    AT_CUDA_CHECK(cudaEventRecord(event_, stream));
+    const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
+    if (C10_UNLIKELY(interp)) {
+      (*interp)->trace_gpu_event_record(
+          reinterpret_cast<uintptr_t>(event_),
+          reinterpret_cast<uintptr_t>(stream.stream())
+      );
+    }
+    was_recorded_ = true;
+  }
+
+  // Note: cudaStreamWaitEvent must be called on the same device as the stream.
+  // The event has no actual GPU resources associated with it.
+  void block(const CUDAStream& stream) {
+    if (is_created_) {
+      CUDAGuard guard(stream.device_index());
+      AT_CUDA_CHECK(cudaStreamWaitEvent(stream, event_, 0));
+      const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
+      if (C10_UNLIKELY(interp)) {
+        (*interp)->trace_gpu_event_wait(
+            reinterpret_cast<uintptr_t>(event_),
+            reinterpret_cast<uintptr_t>(stream.stream())
+        );
+      }
+    }
+  }
+
+  // Note: cudaEventElapsedTime can be safely called from any device
+  float elapsed_time(const CUDAEvent& other) const {
+    TORCH_CHECK(is_created_ && other.isCreated(),
+      "Both events must be recorded before calculating elapsed time.");
+    float time_ms = 0;
+    // raise cudaErrorNotReady if either event is recorded but not yet completed
+    AT_CUDA_CHECK(cudaEventElapsedTime(&time_ms, event_, other.event_));
+    return time_ms;
+  }
+
+  // Note: cudaEventSynchronize can be safely called from any device
+  void synchronize() const {
+    if (is_created_) {
+      const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
+      if (C10_UNLIKELY(interp)) {
+          (*interp)->trace_gpu_event_synchronization(reinterpret_cast<uintptr_t>(event_));
+      }
+      AT_CUDA_CHECK(cudaEventSynchronize(event_));
+    }
+  }
+
+  // Note: cudaIpcGetEventHandle must be called on the same device as the event
+  void ipc_handle(cudaIpcEventHandle_t * handle) {
+      if (!is_created_) {
+        // this CUDAEvent object was initially constructed from flags but event_
+        // is not created yet.
+        createEvent(getCurrentCUDAStream().device_index());
+      }
+      CUDAGuard guard(device_index_);
+      AT_CUDA_CHECK(cudaIpcGetEventHandle(handle, event_));
+  }
+
+private:
+  unsigned int flags_ = cudaEventDisableTiming;
+  bool is_created_ = false;
+  bool was_recorded_ = false;
+  DeviceIndex device_index_ = -1;
+  cudaEvent_t event_{};
+
+  void createEvent(DeviceIndex device_index) {
+    device_index_ = device_index;
+    CUDAGuard guard(device_index_);
+    AT_CUDA_CHECK(cudaEventCreateWithFlags(&event_, flags_));
+    const c10::impl::PyInterpreter* interp = c10::impl::GPUTrace::get_trace();
+    if (C10_UNLIKELY(interp)) {
+      (*interp)->trace_gpu_event_creation(reinterpret_cast<uintptr_t>(event_));
+    }
+    is_created_ = true;
+  }
+
+  void moveHelper(CUDAEvent&& other) {
+    std::swap(flags_, other.flags_);
+    std::swap(is_created_, other.is_created_);
+    std::swap(was_recorded_, other.was_recorded_);
+    std::swap(device_index_, other.device_index_);
+    std::swap(event_, other.event_);
+  }
+};
+
+} // namespace at::cuda

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/cuda/CUDASparse.h

@@ -0,0 +1,76 @@
+#pragma once
+
+#include <ATen/cuda/CUDAContext.h>
+#if defined(USE_ROCM)
+#include <hipsparse/hipsparse-version.h>
+#define HIPSPARSE_VERSION ((hipsparseVersionMajor*100000) + (hipsparseVersionMinor*100) + hipsparseVersionPatch)
+#endif
+
+// cuSparse Generic API added in CUDA 10.1
+// Windows support added in CUDA 11.0
+#if defined(CUDART_VERSION) && defined(CUSPARSE_VERSION) && ((CUSPARSE_VERSION >= 10300) || (CUSPARSE_VERSION >= 11000 && defined(_WIN32)))
+#define AT_USE_CUSPARSE_GENERIC_API() 1
+#else
+#define AT_USE_CUSPARSE_GENERIC_API() 0
+#endif
+
+// cuSparse Generic API descriptor pointers were changed to const in CUDA 12.0
+#if defined(CUDART_VERSION) && defined(CUSPARSE_VERSION) && \
+    (CUSPARSE_VERSION < 12000)
+#define AT_USE_CUSPARSE_NON_CONST_DESCRIPTORS() 1
+#else
+#define AT_USE_CUSPARSE_NON_CONST_DESCRIPTORS() 0
+#endif
+
+#if defined(CUDART_VERSION) && defined(CUSPARSE_VERSION) && \
+    (CUSPARSE_VERSION >= 12000)
+#define AT_USE_CUSPARSE_CONST_DESCRIPTORS() 1
+#else
+#define AT_USE_CUSPARSE_CONST_DESCRIPTORS() 0
+#endif
+
+#if defined(USE_ROCM)
+// hipSparse const API added in v2.4.0
+#if HIPSPARSE_VERSION >= 200400
+#define AT_USE_HIPSPARSE_CONST_DESCRIPTORS() 1
+#define AT_USE_HIPSPARSE_NON_CONST_DESCRIPTORS() 0
+#define AT_USE_HIPSPARSE_GENERIC_API() 1
+#else
+#define AT_USE_HIPSPARSE_CONST_DESCRIPTORS() 0
+#define AT_USE_HIPSPARSE_NON_CONST_DESCRIPTORS() 1
+#define AT_USE_HIPSPARSE_GENERIC_API() 1
+#endif
+#else // USE_ROCM
+#define AT_USE_HIPSPARSE_CONST_DESCRIPTORS() 0
+#define AT_USE_HIPSPARSE_NON_CONST_DESCRIPTORS() 0
+#define AT_USE_HIPSPARSE_GENERIC_API() 0
+#endif // USE_ROCM
+
+// cuSparse Generic API spsv function was added in CUDA 11.3.0
+#if defined(CUDART_VERSION) && defined(CUSPARSE_VERSION) && (CUSPARSE_VERSION >= 11500)
+#define AT_USE_CUSPARSE_GENERIC_SPSV() 1
+#else
+#define AT_USE_CUSPARSE_GENERIC_SPSV() 0
+#endif
+
+// cuSparse Generic API spsm function was added in CUDA 11.3.1
+#if defined(CUDART_VERSION) && defined(CUSPARSE_VERSION) && (CUSPARSE_VERSION >= 11600)
+#define AT_USE_CUSPARSE_GENERIC_SPSM() 1
+#else
+#define AT_USE_CUSPARSE_GENERIC_SPSM() 0
+#endif
+
+// cuSparse Generic API sddmm function was added in CUDA 11.2.1 (cuSparse version 11400)
+#if defined(CUDART_VERSION) && defined(CUSPARSE_VERSION) && (CUSPARSE_VERSION >= 11400)
+#define AT_USE_CUSPARSE_GENERIC_SDDMM() 1
+#else
+#define AT_USE_CUSPARSE_GENERIC_SDDMM() 0
+#endif
+
+// BSR triangular solve functions were added in hipSPARSE 1.11.2 (ROCm 4.5.0)
+#if defined(CUDART_VERSION) ||                            \
+      (defined(USE_ROCM) && ROCM_VERSION >= 40500 )
+#define AT_USE_HIPSPARSE_TRIANGULAR_SOLVE() 1
+#else
+#define AT_USE_HIPSPARSE_TRIANGULAR_SOLVE() 0
+#endif

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/cuda/CUDASparseDescriptors.h

@@ -0,0 +1,290 @@
+#pragma once
+
+#include <ATen/Tensor.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <ATen/cuda/CUDASparse.h>
+
+#include <c10/core/ScalarType.h>
+
+#if defined(USE_ROCM)
+#include <type_traits>
+#endif
+
+namespace at::cuda::sparse {
+
+template <typename T, cusparseStatus_t (*destructor)(T*)>
+struct CuSparseDescriptorDeleter {
+  void operator()(T* x) {
+    if (x != nullptr) {
+      TORCH_CUDASPARSE_CHECK(destructor(x));
+    }
+  }
+};
+
+template <typename T, cusparseStatus_t (*destructor)(T*)>
+class CuSparseDescriptor {
+ public:
+  T* descriptor() const {
+    return descriptor_.get();
+  }
+  T* descriptor() {
+    return descriptor_.get();
+  }
+
+ protected:
+  std::unique_ptr<T, CuSparseDescriptorDeleter<T, destructor>> descriptor_;
+};
+
+#if AT_USE_CUSPARSE_CONST_DESCRIPTORS() || AT_USE_HIPSPARSE_CONST_DESCRIPTORS()
+template <typename T, cusparseStatus_t (*destructor)(const T*)>
+struct ConstCuSparseDescriptorDeleter {
+  void operator()(T* x) {
+    if (x != nullptr) {
+      TORCH_CUDASPARSE_CHECK(destructor(x));
+    }
+  }
+};
+
+template <typename T, cusparseStatus_t (*destructor)(const T*)>
+class ConstCuSparseDescriptor {
+ public:
+  T* descriptor() const {
+    return descriptor_.get();
+  }
+  T* descriptor() {
+    return descriptor_.get();
+  }
+
+ protected:
+  std::unique_ptr<T, ConstCuSparseDescriptorDeleter<T, destructor>> descriptor_;
+};
+#endif // AT_USE_CUSPARSE_CONST_DESCRIPTORS || AT_USE_HIPSPARSE_CONST_DESCRIPTORS
+
+#if defined(USE_ROCM)
+using cusparseMatDescr = std::remove_pointer<hipsparseMatDescr_t>::type;
+using cusparseDnMatDescr = std::remove_pointer<hipsparseDnMatDescr_t>::type;
+using cusparseDnVecDescr = std::remove_pointer<hipsparseDnVecDescr_t>::type;
+using cusparseSpMatDescr = std::remove_pointer<hipsparseSpMatDescr_t>::type;
+using cusparseSpMatDescr = std::remove_pointer<hipsparseSpMatDescr_t>::type;
+using cusparseSpGEMMDescr = std::remove_pointer<hipsparseSpGEMMDescr_t>::type;
+#if AT_USE_HIPSPARSE_TRIANGULAR_SOLVE()
+using bsrsv2Info = std::remove_pointer<bsrsv2Info_t>::type;
+using bsrsm2Info = std::remove_pointer<bsrsm2Info_t>::type;
+#endif
+#endif
+
+// NOTE: This is only needed for CUDA 11 and earlier, since CUDA 12 introduced
+// API for const descriptors
+cusparseStatus_t destroyConstDnMat(const cusparseDnMatDescr* dnMatDescr);
+
+class TORCH_CUDA_CPP_API CuSparseMatDescriptor
+    : public CuSparseDescriptor<cusparseMatDescr, &cusparseDestroyMatDescr> {
+ public:
+  CuSparseMatDescriptor() {
+    cusparseMatDescr_t raw_descriptor;
+    TORCH_CUDASPARSE_CHECK(cusparseCreateMatDescr(&raw_descriptor));
+    descriptor_.reset(raw_descriptor);
+  }
+
+  CuSparseMatDescriptor(bool upper, bool unit) {
+    cusparseFillMode_t fill_mode =
+        upper ? CUSPARSE_FILL_MODE_UPPER : CUSPARSE_FILL_MODE_LOWER;
+    cusparseDiagType_t diag_type =
+        unit ? CUSPARSE_DIAG_TYPE_UNIT : CUSPARSE_DIAG_TYPE_NON_UNIT;
+    cusparseMatDescr_t raw_descriptor;
+    TORCH_CUDASPARSE_CHECK(cusparseCreateMatDescr(&raw_descriptor));
+    TORCH_CUDASPARSE_CHECK(cusparseSetMatFillMode(raw_descriptor, fill_mode));
+    TORCH_CUDASPARSE_CHECK(cusparseSetMatDiagType(raw_descriptor, diag_type));
+    descriptor_.reset(raw_descriptor);
+  }
+};
+
+#if AT_USE_HIPSPARSE_TRIANGULAR_SOLVE()
+
+class TORCH_CUDA_CPP_API CuSparseBsrsv2Info
+    : public CuSparseDescriptor<bsrsv2Info, &cusparseDestroyBsrsv2Info> {
+ public:
+  CuSparseBsrsv2Info() {
+    bsrsv2Info_t raw_descriptor;
+    TORCH_CUDASPARSE_CHECK(cusparseCreateBsrsv2Info(&raw_descriptor));
+    descriptor_.reset(raw_descriptor);
+  }
+};
+
+class TORCH_CUDA_CPP_API CuSparseBsrsm2Info
+    : public CuSparseDescriptor<bsrsm2Info, &cusparseDestroyBsrsm2Info> {
+ public:
+  CuSparseBsrsm2Info() {
+    bsrsm2Info_t raw_descriptor;
+    TORCH_CUDASPARSE_CHECK(cusparseCreateBsrsm2Info(&raw_descriptor));
+    descriptor_.reset(raw_descriptor);
+  }
+};
+
+#endif // AT_USE_HIPSPARSE_TRIANGULAR_SOLVE
+
+#if AT_USE_CUSPARSE_GENERIC_API() || AT_USE_HIPSPARSE_GENERIC_API()
+
+cusparseIndexType_t getCuSparseIndexType(const c10::ScalarType& scalar_type);
+
+#if AT_USE_CUSPARSE_NON_CONST_DESCRIPTORS() || AT_USE_HIPSPARSE_NON_CONST_DESCRIPTORS()
+class TORCH_CUDA_CPP_API CuSparseDnMatDescriptor
+    : public CuSparseDescriptor<cusparseDnMatDescr, &cusparseDestroyDnMat> {
+ public:
+  explicit CuSparseDnMatDescriptor(const Tensor& input, int64_t batch_offset = -1);
+};
+
+class TORCH_CUDA_CPP_API CuSparseConstDnMatDescriptor
+    : public CuSparseDescriptor<const cusparseDnMatDescr, &destroyConstDnMat> {
+ public:
+  explicit CuSparseConstDnMatDescriptor(const Tensor& input, int64_t batch_offset = -1);
+  cusparseDnMatDescr* unsafe_mutable_descriptor() const {
+    return const_cast<cusparseDnMatDescr*>(descriptor());
+  }
+  cusparseDnMatDescr* unsafe_mutable_descriptor() {
+    return const_cast<cusparseDnMatDescr*>(descriptor());
+  }
+};
+
+class TORCH_CUDA_CPP_API CuSparseDnVecDescriptor
+    : public CuSparseDescriptor<cusparseDnVecDescr, &cusparseDestroyDnVec> {
+ public:
+  explicit CuSparseDnVecDescriptor(const Tensor& input);
+};
+
+class TORCH_CUDA_CPP_API CuSparseSpMatDescriptor
+    : public CuSparseDescriptor<cusparseSpMatDescr, &cusparseDestroySpMat> {};
+
+#elif AT_USE_CUSPARSE_CONST_DESCRIPTORS() || AT_USE_HIPSPARSE_CONST_DESCRIPTORS()
+  class TORCH_CUDA_CPP_API CuSparseDnMatDescriptor
+      : public ConstCuSparseDescriptor<
+            cusparseDnMatDescr,
+            &cusparseDestroyDnMat> {
+   public:
+    explicit CuSparseDnMatDescriptor(
+        const Tensor& input,
+        int64_t batch_offset = -1);
+  };
+
+  class TORCH_CUDA_CPP_API CuSparseConstDnMatDescriptor
+      : public ConstCuSparseDescriptor<
+            const cusparseDnMatDescr,
+            &destroyConstDnMat> {
+   public:
+    explicit CuSparseConstDnMatDescriptor(
+        const Tensor& input,
+        int64_t batch_offset = -1);
+  cusparseDnMatDescr* unsafe_mutable_descriptor() const {
+    return const_cast<cusparseDnMatDescr*>(descriptor());
+  }
+  cusparseDnMatDescr* unsafe_mutable_descriptor() {
+    return const_cast<cusparseDnMatDescr*>(descriptor());
+  }
+  };
+
+  class TORCH_CUDA_CPP_API CuSparseDnVecDescriptor
+      : public ConstCuSparseDescriptor<
+            cusparseDnVecDescr,
+            &cusparseDestroyDnVec> {
+   public:
+    explicit CuSparseDnVecDescriptor(const Tensor& input);
+  };
+
+  class TORCH_CUDA_CPP_API CuSparseSpMatDescriptor
+      : public ConstCuSparseDescriptor<
+            cusparseSpMatDescr,
+            &cusparseDestroySpMat> {};
+#endif // AT_USE_CUSPARSE_CONST_DESCRIPTORS() || AT_USE_HIPSPARSE_CONST_DESCRIPTORS()
+
+class TORCH_CUDA_CPP_API CuSparseSpMatCsrDescriptor
+    : public CuSparseSpMatDescriptor {
+ public:
+  explicit CuSparseSpMatCsrDescriptor(const Tensor& input, int64_t batch_offset = -1);
+
+  std::tuple<int64_t, int64_t, int64_t> get_size() {
+    int64_t rows, cols, nnz;
+    TORCH_CUDASPARSE_CHECK(cusparseSpMatGetSize(
+        this->descriptor(),
+        &rows,
+        &cols,
+        &nnz));
+    return std::make_tuple(rows, cols, nnz);
+  }
+
+  void set_tensor(const Tensor& input) {
+    auto crow_indices = input.crow_indices();
+    auto col_indices = input.col_indices();
+    auto values = input.values();
+
+    TORCH_INTERNAL_ASSERT_DEBUG_ONLY(crow_indices.is_contiguous());
+    TORCH_INTERNAL_ASSERT_DEBUG_ONLY(col_indices.is_contiguous());
+    TORCH_INTERNAL_ASSERT_DEBUG_ONLY(values.is_contiguous());
+    TORCH_CUDASPARSE_CHECK(cusparseCsrSetPointers(
+        this->descriptor(),
+        crow_indices.data_ptr(),
+        col_indices.data_ptr(),
+        values.data_ptr()));
+  }
+
+#if AT_USE_CUSPARSE_GENERIC_SPSV()
+  void set_mat_fill_mode(bool upper) {
+    cusparseFillMode_t fill_mode =
+        upper ? CUSPARSE_FILL_MODE_UPPER : CUSPARSE_FILL_MODE_LOWER;
+    TORCH_CUDASPARSE_CHECK(cusparseSpMatSetAttribute(
+        this->descriptor(),
+        CUSPARSE_SPMAT_FILL_MODE,
+        &fill_mode,
+        sizeof(fill_mode)));
+  }
+
+  void set_mat_diag_type(bool unit) {
+    cusparseDiagType_t diag_type =
+        unit ? CUSPARSE_DIAG_TYPE_UNIT : CUSPARSE_DIAG_TYPE_NON_UNIT;
+    TORCH_CUDASPARSE_CHECK(cusparseSpMatSetAttribute(
+        this->descriptor(),
+        CUSPARSE_SPMAT_DIAG_TYPE,
+        &diag_type,
+        sizeof(diag_type)));
+  }
+#endif
+};
+
+#if AT_USE_CUSPARSE_GENERIC_SPSV()
+class TORCH_CUDA_CPP_API CuSparseSpSVDescriptor
+    : public CuSparseDescriptor<cusparseSpSVDescr, &cusparseSpSV_destroyDescr> {
+ public:
+  CuSparseSpSVDescriptor() {
+    cusparseSpSVDescr_t raw_descriptor;
+    TORCH_CUDASPARSE_CHECK(cusparseSpSV_createDescr(&raw_descriptor));
+    descriptor_.reset(raw_descriptor);
+  }
+};
+#endif
+
+#if AT_USE_CUSPARSE_GENERIC_SPSM()
+class TORCH_CUDA_CPP_API CuSparseSpSMDescriptor
+    : public CuSparseDescriptor<cusparseSpSMDescr, &cusparseSpSM_destroyDescr> {
+ public:
+  CuSparseSpSMDescriptor() {
+    cusparseSpSMDescr_t raw_descriptor;
+    TORCH_CUDASPARSE_CHECK(cusparseSpSM_createDescr(&raw_descriptor));
+    descriptor_.reset(raw_descriptor);
+  }
+};
+#endif
+
+#if (defined(USE_ROCM) && ROCM_VERSION >= 50200) || !defined(USE_ROCM)
+class TORCH_CUDA_CPP_API CuSparseSpGEMMDescriptor
+    : public CuSparseDescriptor<cusparseSpGEMMDescr, &cusparseSpGEMM_destroyDescr> {
+ public:
+  CuSparseSpGEMMDescriptor() {
+    cusparseSpGEMMDescr_t raw_descriptor;
+    TORCH_CUDASPARSE_CHECK(cusparseSpGEMM_createDescr(&raw_descriptor));
+    descriptor_.reset(raw_descriptor);
+  }
+};
+#endif
+
+#endif // AT_USE_CUSPARSE_GENERIC_API() || AT_USE_HIPSPARSE_GENERIC_API()
+
+} // namespace at::cuda::sparse

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/cuda/CUDAUtils.h

@@ -0,0 +1,20 @@
+#pragma once
+
+#include <ATen/cuda/CUDAContext.h>
+
+namespace at::cuda {
+
+// Check if every tensor in a list of tensors matches the current
+// device.
+inline bool check_device(ArrayRef<Tensor> ts) {
+  if (ts.empty()) {
+    return true;
+  }
+  Device curDevice = Device(kCUDA, current_device());
+  for (const Tensor& t : ts) {
+    if (t.device() != curDevice) return false;
+  }
+  return true;
+}
+
+} // namespace at::cuda

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/cuda/CachingHostAllocator.h

@@ -0,0 +1,37 @@
+#pragma once
+
+#include <c10/core/Allocator.h>
+#include <c10/cuda/CUDAStream.h>
+
+namespace at::cuda {
+
+//
+// A caching allocator for CUDA host allocations (pinned memory).
+//
+// This provides a drop-in replacement for THCudaHostAllocator, which re-uses
+// freed pinned (page-locked) memory allocations. This avoids device
+// synchronizations due to cudaFreeHost calls.
+//
+// To ensure correct behavior, THCCachingHostAllocator_recordEvent must be
+// called anytime a pointer from this allocator is used in a cudaMemcpyAsync
+// call between host and device, and passed the corresponding context from the
+// allocation. This is currently invoked by at::native::copy_kernel_cuda.
+//
+// Note that this allocator does not split larger allocations into smaller
+// blocks, unlike the caching device allocator.
+//
+TORCH_CUDA_CPP_API c10::Allocator* getCachingHostAllocator();
+
+// Records an event in the specified stream. The allocation corresponding to the
+// input `ptr`/`ctx` will not be re-used until the event has occurred.
+TORCH_CUDA_CPP_API bool
+CachingHostAllocator_recordEvent(void* ptr, void* ctx, c10::cuda::CUDAStream stream);
+
+// Releases cached pinned memory allocations via cudaHostFree
+TORCH_CUDA_CPP_API void CachingHostAllocator_emptyCache();
+
+inline TORCH_CUDA_CPP_API at::DataPtr HostAlloc(size_t size) {
+  return getCachingHostAllocator()->allocate(size);
+}
+
+} // namespace at::cuda

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/cuda/DeviceUtils.cuh

@@ -0,0 +1,121 @@
+#pragma once
+
+#include <cuda.h>
+#include <c10/util/complex.h>
+#include <c10/util/Half.h>
+
+__device__ __forceinline__ unsigned int ACTIVE_MASK()
+{
+#if !defined(USE_ROCM)
+    return __activemask();
+#else
+// will be ignored anyway
+    return 0xffffffff;
+#endif
+}
+
+__device__ __forceinline__ void WARP_SYNC(unsigned mask = 0xffffffff) {
+#if !defined(USE_ROCM)
+  return __syncwarp(mask);
+#endif
+}
+
+#if defined(USE_ROCM)
+__device__ __forceinline__ unsigned long long int WARP_BALLOT(int predicate)
+{
+return __ballot(predicate);
+}
+#else
+__device__ __forceinline__ unsigned int WARP_BALLOT(int predicate, unsigned int mask = 0xffffffff)
+{
+#if !defined(USE_ROCM)
+    return __ballot_sync(mask, predicate);
+#else
+    return __ballot(predicate);
+#endif
+}
+#endif
+
+template <typename T>
+__device__ __forceinline__ T WARP_SHFL_XOR(T value, int laneMask, int width = warpSize, unsigned int mask = 0xffffffff)
+{
+#if !defined(USE_ROCM)
+    return __shfl_xor_sync(mask, value, laneMask, width);
+#else
+    return __shfl_xor(value, laneMask, width);
+#endif
+}
+
+template <typename T>
+__device__ __forceinline__ T WARP_SHFL(T value, int srcLane, int width = warpSize, unsigned int mask = 0xffffffff)
+{
+#if !defined(USE_ROCM)
+    return __shfl_sync(mask, value, srcLane, width);
+#else
+    return __shfl(value, srcLane, width);
+#endif
+}
+
+template <typename T>
+__device__ __forceinline__ T WARP_SHFL_UP(T value, unsigned int delta, int width = warpSize, unsigned int mask = 0xffffffff)
+{
+#if !defined(USE_ROCM)
+    return __shfl_up_sync(mask, value, delta, width);
+#else
+    return __shfl_up(value, delta, width);
+#endif
+}
+
+template <typename T>
+__device__ __forceinline__ T WARP_SHFL_DOWN(T value, unsigned int delta, int width = warpSize, unsigned int mask = 0xffffffff)
+{
+#if !defined(USE_ROCM)
+    return __shfl_down_sync(mask, value, delta, width);
+#else
+    return __shfl_down(value, delta, width);
+#endif
+}
+
+#if defined(USE_ROCM)
+template<>
+__device__ __forceinline__ int64_t WARP_SHFL_DOWN<int64_t>(int64_t value, unsigned int delta, int width , unsigned int mask)
+{
+  //(HIP doesn't support int64_t). Trick from https://devblogs.nvidia.com/faster-parallel-reductions-kepler/
+  int2 a = *reinterpret_cast<int2*>(&value);
+  a.x = __shfl_down(a.x, delta);
+  a.y = __shfl_down(a.y, delta);
+  return *reinterpret_cast<int64_t*>(&a);
+}
+#endif
+
+template<>
+__device__ __forceinline__ c10::Half WARP_SHFL_DOWN<c10::Half>(c10::Half value, unsigned int delta, int width, unsigned int mask)
+{
+  return c10::Half(WARP_SHFL_DOWN<unsigned short>(value.x, delta, width, mask), c10::Half::from_bits_t{});
+}
+
+template <typename T>
+__device__ __forceinline__ c10::complex<T> WARP_SHFL_DOWN(c10::complex<T> value, unsigned int delta, int width = warpSize, unsigned int mask = 0xffffffff)
+{
+#if !defined(USE_ROCM)
+    return c10::complex<T>(
+        __shfl_down_sync(mask, value.real_, delta, width),
+        __shfl_down_sync(mask, value.imag_, delta, width));
+#else
+    return c10::complex<T>(
+        __shfl_down(value.real_, delta, width),
+        __shfl_down(value.imag_, delta, width));
+#endif
+}
+
+/**
+ * For CC 3.5+, perform a load using __ldg
+ */
+template <typename T>
+__device__ __forceinline__ T doLdg(const T* p) {
+#if __CUDA_ARCH__ >= 350 && !defined(USE_ROCM)
+  return __ldg(p);
+#else
+  return *p;
+#endif
+}

tuning-competition-baseline/.venv/lib/python3.11/site-packages/torch/include/ATen/cuda/EmptyTensor.h

@@ -0,0 +1,44 @@
+#pragma once
+#include <ATen/core/TensorBase.h>
+
+namespace at::detail {
+
+TORCH_CUDA_CPP_API TensorBase empty_cuda(
+    IntArrayRef size,
+    ScalarType dtype,
+    c10::optional<Device> device_opt,
+    c10::optional<c10::MemoryFormat> memory_format_opt);
+
+TORCH_CUDA_CPP_API TensorBase empty_cuda(
+    IntArrayRef size,
+    c10::optional<ScalarType> dtype_opt,
+    c10::optional<Layout> layout_opt,
+    c10::optional<Device> device_opt,
+    c10::optional<bool> pin_memory_opt,
+    c10::optional<c10::MemoryFormat> memory_format_opt);
+
+TORCH_CUDA_CPP_API TensorBase empty_cuda(
+    IntArrayRef size,
+    const TensorOptions &options);
+
+TORCH_CUDA_CPP_API TensorBase empty_strided_cuda(
+    IntArrayRef size,
+    IntArrayRef stride,
+    ScalarType dtype,
+    c10::optional<Device> device_opt);
+
+TORCH_CUDA_CPP_API TensorBase empty_strided_cuda(
+    IntArrayRef size,
+    IntArrayRef stride,
+    c10::optional<ScalarType> dtype_opt,
+    c10::optional<Layout> layout_opt,
+    c10::optional<Device> device_opt,
+    c10::optional<bool> pin_memory_opt);
+
+TORCH_CUDA_CPP_API TensorBase empty_strided_cuda(
+    IntArrayRef size,
+    IntArrayRef stride,
+    const TensorOptions &options);
+
+
+}  // namespace at::detail