Dynamic Padding bug, with a minimal fix!

[kistenklaus]

Here is a minimal example to reproduce the issue:

import torch
class Net(torch.nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, I):
        H = I.size(2) # H is symbolic (dynamic_shape)
        x = torch.nn.functional.pad(I, (0, 0, 0, H), mode="replicate")
        return x

torch.onnx.export(Net(), 
                  (torch.rand(1,1,1,1),), 
                  "net.onnx",
                  dynamo=True,
                  dynamic_shapes={"I" : {2 : torch.export.Dim("H")}},
                  report=True)

Now i'm very much not a python dev, actually have ever done anything in python.
So take anything i say with a grain of salt =^).

But i ran into this issue, because in U-Nets for image denoising we sometimes pad upto an alignment, to ensure that following concat nodes see, two tensors with identical spatial extent, which means that at least in this field dynamic padding is a incredibly commonly used operation.

Fix:

I looked through the error logs a bit (see below), and i think what's goin on is that the
_process_padding function at /onnxscript/blob/main/onnxscript/function_libs/torch_lib/ops/nn.py is missing a splat operator in the dynamic value path, idk if it's actually called this, but good name =^). Maybe iam also wrong here, not a python dev afterall, but this at least fixed the issue for me:

def _process_padding(padding, rank):
    """Convert PyTorch padding for ONNX Pad."""
    assert isinstance(padding, (list, tuple))
    if all(isinstance(pad, int) for pad in padding):
        paddings = padding
        zeros = [0] * (rank * 2 - len(paddings))
        paddings = [*paddings, *zeros]
        paddings = paddings[-2::-2] + paddings[-1::-2]
        return op.Constant(value=ir.tensor(paddings, dtype=ir.DataType.INT64))
    else:
        paddings = []
        for pad in padding:
            if isinstance(pad, int):
                paddings.append(op.Constant(value_ints=[pad]))
            else:
                # Dynamic value
                paddings.append(op.Reshape(pad, [-1]))
        # Create a series of 1d zero tensors
        zero = op.Constant(value_ints=[0])
        zeros = [zero] * (rank * 2 - len(paddings))
        paddings = [*paddings, *zeros]
        # Interleave the padding values
        paddings = paddings[-2::-2] + paddings[-1::-2]
        return op.Concat(*paddings, axis=0) # <- ONLY LINE THAT'S CHANGED!

torch_nn_ops._process_padding = _process_padding

Markdown Report:

PyTorch ONNX Conversion Report

✅ Obtain model graph with `torch.export.export(..., strict=False)`
⚪ Obtain model graph with `torch.export.export(..., strict=True)`
⚪ Obtain model graph with `torch.export._draft_export.draft_export`
✅ Decompose operators for ONNX compatibility
❌ Translate the graph into ONNX
⚪ Run `onnx.checker` on the ONNX model
⚪ Execute the model with ONNX Runtime
⚪ Validate model output accuracy

Error messages

Traceback (most recent call last):

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/onnx/_internal/exporter/_building.py", line 565, in _call_op
    converted_named_inputs = _process_python_sequences(
        op_signature,
    ...<3 lines>...
        self.opset,
    )

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/onnx/_internal/exporter/_building.py", line 415, in _process_python_sequences
    dtype = _determine_input_dtype(param, arg, type_binding)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/onnx/_internal/exporter/_building.py", line 234, in _determine_input_dtype
    raise ValueError(
    ...<3 lines>...
    )

ValueError: Could not determine the dtype for the input 'inputs'. param=inputs: T, arg=([SymbolicTensor(name='anonymous:139994164589296', producer=anonymous_node:139994164995296, index=0), SymbolicTensor(name='anonymous:139994164589296', producer=anonymous_node:139994164995296, index=0), SymbolicTensor(name='anonymous:139994164589456', producer=anonymous_node:139994164994864, index=0), SymbolicTensor(name='anonymous:139994164589776', producer=anonymous_node:139994164994432, index=0), SymbolicTensor(name='anonymous:139994164589296', producer=anonymous_node:139994164995296, index=0), SymbolicTensor(name='anonymous:139994164589296', producer=anonymous_node:139994164995296, index=0), SymbolicTensor(name='anonymous:139994164588816', producer=anonymous_node:139994164995152, index=0), SymbolicTensor(name='anonymous:139994164589616', producer=anonymous_node:139994164994720, index=0)],), param_type_constraint=T=INT16 | UINT16 | INT64 | FLOAT16 | COMPLEX64 | BOOL | STRING | COMPLEX128 | UINT8 | INT32 | DOUBLE | BFLOAT16 | UINT64 | FLOAT | INT8 | UINT32, type_binding={}


The above exception was the direct cause of the following exception:


Traceback (most recent call last):

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/onnx/_internal/exporter/_building.py", line 630, in eval
    outputs = self._call_op(
        op_signature, named_inputs, named_attrs, num_outputs
    )

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/onnx/_internal/exporter/_building.py", line 574, in _call_op
    raise _errors.GraphConstructionError(
    ...<2 lines>...
    ) from e

torch.onnx._internal.exporter._errors.GraphConstructionError: Error processing Python constants for operator '::Concat'. named_inputs={'inputs': ([SymbolicTensor(name='anonymous:139994164589296', producer=anonymous_node:139994164995296, index=0), SymbolicTensor(name='anonymous:139994164589296', producer=anonymous_node:139994164995296, index=0), SymbolicTensor(name='anonymous:139994164589456', producer=anonymous_node:139994164994864, index=0), SymbolicTensor(name='anonymous:139994164589776', producer=anonymous_node:139994164994432, index=0), SymbolicTensor(name='anonymous:139994164589296', producer=anonymous_node:139994164995296, index=0), SymbolicTensor(name='anonymous:139994164589296', producer=anonymous_node:139994164995296, index=0), SymbolicTensor(name='anonymous:139994164588816', producer=anonymous_node:139994164995152, index=0), SymbolicTensor(name='anonymous:139994164589616', producer=anonymous_node:139994164994720, index=0)],)}, named_attrs={'axis': 0}, opset=, op_signature=''::Concat(inputs: T, axis: INT = None) -> (T) where T=INT16 | UINT16 | INT64 | FLOAT16 | COMPLEX64 | BOOL | STRING | COMPLEX128 | UINT8 | INT32 | DOUBLE | BFLOAT16 | UINT64 | FLOAT | INT8 | UINT32.


The above exception was the direct cause of the following exception:


Traceback (most recent call last):

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/onnx/_internal/exporter/_core.py", line 591, in _handle_call_function_node_with_lowering
    outputs = onnx_function(*onnx_args, **onnx_kwargs)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/onnxscript/values.py", line 625, in __call__
    return self.func(*args, **kwargs)
           ~~~~~~~~~^^^^^^^^^^^^^^^^^

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/onnxscript/function_libs/torch_lib/ops/nn.py", line 1519, in aten_pad
    paddings = _process_padding(pad, rank)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/onnxscript/function_libs/torch_lib/ops/nn.py", line 1506, in _process_padding
    return op.Concat(paddings, axis=0)
           ~~~~~~~~~^^^^^^^^^^^^^^^^^^

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/onnxscript/onnx_opset/_impl/opset13.py", line 389, in Concat
    return op(*self._prepare_inputs(schema, *inputs), axis=axis)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/onnxscript/values.py", line 309, in __call__
    return evaluator.default().eval(schema, args, kwargs)
           ~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/onnx/_internal/exporter/_building.py", line 637, in eval
    raise _errors.GraphConstructionError(
        f"Error calling operator '{schema.name}' with args {args} and kwargs {kwargs}."
    ) from e

torch.onnx._internal.exporter._errors.GraphConstructionError: Error calling operator 'Concat' with args ([SymbolicTensor(name='anonymous:139994164589296', producer=anonymous_node:139994164995296, index=0), SymbolicTensor(name='anonymous:139994164589296', producer=anonymous_node:139994164995296, index=0), SymbolicTensor(name='anonymous:139994164589456', producer=anonymous_node:139994164994864, index=0), SymbolicTensor(name='anonymous:139994164589776', producer=anonymous_node:139994164994432, index=0), SymbolicTensor(name='anonymous:139994164589296', producer=anonymous_node:139994164995296, index=0), SymbolicTensor(name='anonymous:139994164589296', producer=anonymous_node:139994164995296, index=0), SymbolicTensor(name='anonymous:139994164588816', producer=anonymous_node:139994164995152, index=0), SymbolicTensor(name='anonymous:139994164589616', producer=anonymous_node:139994164994720, index=0)],) and kwargs {'axis': 0}.


The above exception was the direct cause of the following exception:


Traceback (most recent call last):

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/onnx/_internal/exporter/_core.py", line 790, in _translate_fx_graph
    _handle_call_function_node_with_lowering(
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^
        model,
        ^^^^^^
    ...<6 lines>...
        node_name_to_local_functions=node_name_to_local_functions,
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    )
    ^

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/onnx/_internal/exporter/_core.py", line 593, in _handle_call_function_node_with_lowering
    raise _errors.GraphConstructionError(
        f"Error when calling function '{onnx_function}' with args '{onnx_args}' and kwargs '{onnx_kwargs}'"
    ) from e

torch.onnx._internal.exporter._errors.GraphConstructionError: Error when calling function 'TracedOnnxFunction(<function aten_pad at 0x7f52f0ba9ee0>)' with args '[SymbolicTensor(name='i', type=Tensor(FLOAT), shape=Shape([1, 1, SymbolicDim(s95), 1])), [0, 0, 0, SymbolicTensor(name='sym_size_int_2', type=Tensor(INT64), shape=Shape([]), producer='node_sym_size_int_2', index=0)], 'replicate']' and kwargs '{}'


The above exception was the direct cause of the following exception:


Traceback (most recent call last):

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/onnx/_internal/exporter/_core.py", line 1459, in export
    onnx_program = _exported_program_to_onnx_program(
        decomposed_program, registry=registry
    )

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/onnx/_internal/exporter/_core.py", line 1090, in _exported_program_to_onnx_program
    values = _translate_fx_graph(
        fx_graph,
    ...<4 lines>...
        registry=registry,
    )

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/onnx/_internal/exporter/_core.py", line 816, in _translate_fx_graph
    raise _errors.ConversionError(
        f"Error when translating node {node.format_node()}. See the stack trace for more information."
    ) from e

torch.onnx._internal.exporter._errors.ConversionError: Error when translating node %pad : [num_users=1] = call_function[target=torch.ops.aten.pad.default](args = (%i, [0, 0, 0, %sym_size_int_2], replicate), kwargs = {}). See the stack trace for more information.

Exported program

ExportedProgram:
    class GraphModule(torch.nn.Module):
        def forward(self, i: "f32[1, 1, s95, 1]"):
             # 
            sym_size_int_2: "Sym(s95)" = torch.ops.aten.sym_size.int(i, 2)
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/fuckyou_dynamo.py:47 in forward, code: x = torch.nn.functional.pad(I, (0, 0, 0, H), mode="replicate")
            pad: "f32[1, 1, 2*s95, 1]" = torch.ops.aten.pad.default(i, [0, 0, 0, sym_size_int_2], 'replicate');  i = sym_size_int_2 = None
            return (pad,)
            
Graph signature: 
    # inputs
    i: USER_INPUT
    
    # outputs
    pad: USER_OUTPUT
    
Range constraints: {s95: VR[0, int_oo]}

Analysis

PyTorch ONNX Conversion Analysis

Model Information

The model has 0 parameters and 0 buffers (non-trainable parameters).
Number of parameters per dtype:

defaultdict(<class 'int'>, {})

Number of buffers per dtype:

defaultdict(<class 'int'>, {})

Inputs:

• i: TensorMetadata(shape=torch.Size([1, 1, s95, 1]), dtype=torch.float32, requires_grad=False, stride=(s95, s95, 1, 1), memory_format=torch.channels_last, is_quantized=False, qparams={})

Outputs:

• pad: TensorMetadata(shape=torch.Size([1, 1, 2*s95, 1]), dtype=torch.float32, requires_grad=False, stride=(2*s95, 2*s95, 1, 1), memory_format=torch.channels_last, is_quantized=False, qparams={})

The FX graph has 4 nodes in total. Number of FX nodes per op:

• placeholder: 1
• call_function: 2
• output: 1

Of the call_function nodes, the counts of operators used are:

• aten.sym_size.int: 1
• aten.pad.default: 1

ONNX Conversion Information

All operators in the model have registered ONNX decompositions.

Decomposition comparison

Ops exist only in the ExportedProgram before decomposition: []

Ops exist only in the ExportedProgram after decomposition: []

[kistenklaus]

thanks for the quick merge <3.

[kistenklaus]

Just some food for thought, exporting this model still screams a bunch of warning, but actually now produces the correct result.

class Net(nn.Module):
    def __init__(self):
        super().__init__()
        pm = 'zeros'
        self.enc0 = nn.Conv2d(INPUT_CHANNELS_COUNT, 32, 3, padding='same', padding_mode=pm)
        self.enc1 = nn.Conv2d(32, 48, 3, padding='same', padding_mode=pm)
        self.enc2 = nn.Conv2d(48, 64, 3, padding='same', padding_mode=pm)
        self.enc3 = nn.Conv2d(64, 80, 3, padding='same', padding_mode=pm)
        self.enc4 = nn.Conv2d(80, 112, 3, padding='same', padding_mode=pm)
        self.enc5 = nn.Conv2d(112, 112, 3, padding='same', padding_mode=pm)

        # self.extr = nn.Conv2d(32, 32, 3, padding='same', padding_mode=pm)
        
        self.dec0 = nn.Conv2d(112+112, 112, 3, padding='same', padding_mode=pm)
        self.con0 = nn.Conv2d(112, 112, 3, padding='same', padding_mode=pm)
        self.dec1 = nn.Conv2d(112+80, 80, 3, padding='same', padding_mode=pm)
        self.con1 = nn.Conv2d(80, 80, 3, padding='same', padding_mode=pm)
        self.dec2 = nn.Conv2d(80+64, 64, 3, padding='same', padding_mode=pm)
        self.con2 = nn.Conv2d(64, 64, 3, padding='same', padding_mode=pm)
        self.dec3 = nn.Conv2d(64+48, 48, 3, padding='same', padding_mode=pm)
        self.con3 = nn.Conv2d(48, 48, 3, padding='same', padding_mode=pm)
        self.dec4 = nn.Conv2d(48+32, 16, 3, padding='same', padding_mode=pm)
        self.con4 = nn.Conv2d(16, 16, 3, padding='same', padding_mode=pm)
        self.dec5 = nn.Conv2d(16+INPUT_CHANNELS_COUNT, 12, 3, padding='same', padding_mode=pm)
        self.con5 = nn.Conv2d(12, 12, 3, padding='same', padding_mode=pm)
        
        self.pool = nn.MaxPool2d(2, 2)
        self.upsample = nn.Upsample(scale_factor=2, mode='nearest')

    def forward(self, I):

        alignment = 64  # ensure even H/W so pool+upsample align perfectly
        H, W = I.size(2), I.size(3)
        pad_w = (alignment - (W % alignment)) % alignment
        pad_h = (alignment - (H % alignment)) % alignment
        I_aligned = F.pad(I, (0, pad_w, 0, pad_h), mode="replicate")

        extr = F.relu(self.enc0(I_aligned))
        x_128 = self.pool(extr) 
        x_64  = self.pool(F.relu(self.enc1(x_128)))
        x_32  = self.pool(F.relu(self.enc2(x_64)))
        x_16  = self.pool(F.relu(self.enc3(x_32)))
        x_8   = self.pool(F.relu(self.enc4(x_16)))
        x_4   = self.pool(F.relu(self.enc5(x_8)))
        x     = F.relu(self.dec0(torch.cat([self.upsample(x_4), x_8],   1)))
        x     = F.relu(self.con0(x))
        x     = F.relu(self.dec1(torch.cat([self.upsample(x),   x_16],  1)))
        x     = F.relu(self.con1(x))
        x     = F.relu(self.dec2(torch.cat([self.upsample(x),   x_32],  1)))
        x     = F.relu(self.con2(x))
        x     = F.relu(self.dec3(torch.cat([self.upsample(x),   x_64],  1)))
        x     = F.relu(self.con3(x))
        x     = F.relu(self.dec4(torch.cat([self.upsample(x),   x_128], 1)))
        x     = F.relu(self.con4(x))
        x     = F.relu(self.dec5(torch.cat([self.upsample(x),   I_aligned],  1)))
        x     = F.relu(self.con5(x))

        x = x[:, :, :H, :W]
        return x

Iam not sure if this is because of onnxscript, i think it's a issue in the constraint solver of torch, but no idea, if that can in any way be related to onnx script.

torch.onnx.export(
        Net(), 
        (torch.randn(1,INPUT_CHANNELS_COUNT, 64,64),),
        "net.onnx",
        dynamo=True,
        export_params=True,
        input_names=["input"],
        output_names=["output"],
        dynamic_shapes={"I": {2 : torch.export.Dim("H", min=64, max=128), 3 : torch.export.Dim("W", min=64, max=128)}},
        report=True,
        )

With torch.export.Dim.Auto it works.

[justinchuby]

Could you post the warnings?

[kistenklaus]

yep gimmi a sec

[kistenklaus]

Report:

PyTorch ONNX Conversion Report

❌ Obtain model graph with `torch.export.export(..., strict=False)`
❌ Obtain model graph with `torch.export.export(..., strict=True)`
✅ Obtain model graph with `torch.export._draft_export.draft_export`
✅ Decompose operators for ONNX compatibility
✅ Translate the graph into ONNX
⚪ Run `onnx.checker` on the ONNX model
⚪ Execute the model with ONNX Runtime
⚪ Validate model output accuracy

Error messages

# ⚠️ Errors from strategy 'TorchExportNonStrictStrategy': -----------------------

Traceback (most recent call last):

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 1798, in _export_to_aten_ir_make_fx
    produce_guards_callback(gm)
    ~~~~~~~~~~~~~~~~~~~~~~~^^^^

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 1944, in _produce_guards_callback
    return produce_guards_and_solve_constraints(
        fake_mode=fake_mode,
    ...<4 lines>...
        _is_torch_jit_trace=_is_torch_jit_trace,
    )

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_export/non_strict_utils.py", line 549, in produce_guards_and_solve_constraints
    raise constraint_violation_error

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_export/non_strict_utils.py", line 514, in produce_guards_and_solve_constraints
    shape_env.produce_guards(
    ~~~~~~~~~~~~~~~~~~~~~~~~^
        placeholders,
        ^^^^^^^^^^^^^
    ...<3 lines>...
        ignore_static=False,
        ^^^^^^^^^^^^^^^^^^^^
    )
    ^

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/fx/experimental/symbolic_shapes.py", line 5200, in produce_guards
    return self.produce_guards_verbose(*args, **kwargs, langs=("python",))[0].exprs
           ~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/fx/experimental/symbolic_shapes.py", line 5932, in produce_guards_verbose
    raise ConstraintViolationError(
    ...<3 lines>...
    )

torch.fx.experimental.symbolic_shapes.ConstraintViolationError: Constraints violated (H, W)! For more information, run with TORCH_LOGS="+dynamic".
  - Not all values of H = L['I'].size()[2] in the specified range 64 <= H <= 128 satisfy the generated guard (((-2) + L['I'].size()[2] + ((64 + (-1)*(L['I'].size()[2] % 64)) % 64)) % 2) == 0.
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard (((-2) + L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) % 2) == 0.
  - Not all values of H = L['I'].size()[2] in the specified range 64 <= H <= 128 satisfy the generated guard (((-2) + ((L['I'].size()[2] + ((64 + (-1)*(L['I'].size()[2] % 64)) % 64)) // 2)) % 2) == 0.
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard (((-2) + ((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 2)) % 2) == 0.
  - Not all values of H = L['I'].size()[2] in the specified range 64 <= H <= 128 satisfy the generated guard (((-2) + ((L['I'].size()[2] + ((64 + (-1)*(L['I'].size()[2] % 64)) % 64)) // 4)) % 2) == 0.
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard (((-2) + ((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 4)) % 2) == 0.
  - Not all values of H = L['I'].size()[2] in the specified range 64 <= H <= 128 satisfy the generated guard (((-2) + ((L['I'].size()[2] + ((64 + (-1)*(L['I'].size()[2] % 64)) % 64)) // 8)) % 2) == 0.
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard (((-2) + ((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 8)) % 2) == 0.
  - Not all values of H = L['I'].size()[2] in the specified range 64 <= H <= 128 satisfy the generated guard (((-2) + ((L['I'].size()[2] + ((64 + (-1)*(L['I'].size()[2] % 64)) % 64)) // 16)) % 2) == 0.
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard (((-2) + ((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 16)) % 2) == 0.
  - Not all values of H = L['I'].size()[2] in the specified range 64 <= H <= 128 satisfy the generated guard (((-2) + ((L['I'].size()[2] + ((64 + (-1)*(L['I'].size()[2] % 64)) % 64)) // 32)) % 2) == 0.
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard (((-2) + ((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 32)) % 2) == 0.
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard max(1, (L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64) <= 1.
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard max(1, 2*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)) == (2*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)).
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard max(1, 4*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)) == (4*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)).
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard max(1, 8*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)) == (8*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)).
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard max(1, 16*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)) == (16*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)).
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard max(1, 32*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)) == (32*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)).
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard max(1, 64*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)) == (64*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)).
Suggested fixes:
  _H = Dim('_H', min=1, max=2)
  _W = Dim('_W', min=1, max=2)
  H = 64*_H
  W = 64*_W


During handling of the above exception, another exception occurred:


Traceback (most recent call last):

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/onnx/_internal/exporter/_capture_strategies.py", line 210, in _capture
    return torch.export.export(
           ~~~~~~~~~~~~~~~~~~~^
        model,
        ^^^^^^
    ...<3 lines>...
        strict=False,
        ^^^^^^^^^^^^^
    )
    ^

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/__init__.py", line 319, in export
    raise e

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/__init__.py", line 286, in export
    return _export(
        mod,
    ...<5 lines>...
        pre_dispatch=True,
    )

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 1164, in wrapper
    raise e

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 1130, in wrapper
    ep = fn(*args, **kwargs)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/exported_program.py", line 123, in wrapper
    return fn(*args, **kwargs)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 2176, in _export
    ep = _export_for_training(
        mod,
    ...<4 lines>...
        preserve_module_call_signature=preserve_module_call_signature,
    )

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 1164, in wrapper
    raise e

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 1130, in wrapper
    ep = fn(*args, **kwargs)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/exported_program.py", line 123, in wrapper
    return fn(*args, **kwargs)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 2037, in _export_for_training
    export_artifact = export_func(
        mod=mod,
    ...<7 lines>...
        _to_aten_func=_export_to_aten_ir_make_fx,
    )

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 1979, in _non_strict_export
    aten_export_artifact = _to_aten_func(  # type: ignore[operator]
        patched_mod,
    ...<5 lines>...
        transform=_tuplify_outputs,
    )

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 1800, in _export_to_aten_ir_make_fx
    raise UserError(UserErrorType.CONSTRAINT_VIOLATION, str(e))  # noqa: B904
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

torch._dynamo.exc.UserError: Constraints violated (H, W)! For more information, run with TORCH_LOGS="+dynamic".
  - Not all values of H = L['I'].size()[2] in the specified range 64 <= H <= 128 satisfy the generated guard (((-2) + L['I'].size()[2] + ((64 + (-1)*(L['I'].size()[2] % 64)) % 64)) % 2) == 0.
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard (((-2) + L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) % 2) == 0.
  - Not all values of H = L['I'].size()[2] in the specified range 64 <= H <= 128 satisfy the generated guard (((-2) + ((L['I'].size()[2] + ((64 + (-1)*(L['I'].size()[2] % 64)) % 64)) // 2)) % 2) == 0.
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard (((-2) + ((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 2)) % 2) == 0.
  - Not all values of H = L['I'].size()[2] in the specified range 64 <= H <= 128 satisfy the generated guard (((-2) + ((L['I'].size()[2] + ((64 + (-1)*(L['I'].size()[2] % 64)) % 64)) // 4)) % 2) == 0.
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard (((-2) + ((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 4)) % 2) == 0.
  - Not all values of H = L['I'].size()[2] in the specified range 64 <= H <= 128 satisfy the generated guard (((-2) + ((L['I'].size()[2] + ((64 + (-1)*(L['I'].size()[2] % 64)) % 64)) // 8)) % 2) == 0.
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard (((-2) + ((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 8)) % 2) == 0.
  - Not all values of H = L['I'].size()[2] in the specified range 64 <= H <= 128 satisfy the generated guard (((-2) + ((L['I'].size()[2] + ((64 + (-1)*(L['I'].size()[2] % 64)) % 64)) // 16)) % 2) == 0.
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard (((-2) + ((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 16)) % 2) == 0.
  - Not all values of H = L['I'].size()[2] in the specified range 64 <= H <= 128 satisfy the generated guard (((-2) + ((L['I'].size()[2] + ((64 + (-1)*(L['I'].size()[2] % 64)) % 64)) // 32)) % 2) == 0.
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard (((-2) + ((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 32)) % 2) == 0.
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard max(1, (L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64) <= 1.
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard max(1, 2*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)) == (2*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)).
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard max(1, 4*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)) == (4*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)).
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard max(1, 8*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)) == (8*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)).
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard max(1, 16*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)) == (16*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)).
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard max(1, 32*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)) == (32*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)).
  - Not all values of W = L['I'].size()[3] in the specified range 64 <= W <= 128 satisfy the generated guard max(1, 64*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)) == (64*((L['I'].size()[3] + ((64 + (-1)*(L['I'].size()[3] % 64)) % 64)) // 64)).
Suggested fixes:
  _H = Dim('_H', min=1, max=2)
  _W = Dim('_W', min=1, max=2)
  H = 64*_H
  W = 64*_W

The error above occurred when calling torch.export.export. If you would like to view some more information about this error, and get a list of all other errors that may occur in your export call, you can replace your `export()` call with `draft_export()`.


During handling of the above exception, another exception occurred:


Traceback (most recent call last):

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/onnx/_internal/exporter/_capture_strategies.py", line 118, in __call__
    exported_program = self._capture(model, args, kwargs, dynamic_shapes)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/onnx/_internal/exporter/_capture_strategies.py", line 226, in _capture
    return torch.export.export(
           ~~~~~~~~~~~~~~~~~~~^
        model, args, kwargs=kwargs, dynamic_shapes=new_shapes, strict=False
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    )
    ^

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/__init__.py", line 319, in export
    raise e

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/__init__.py", line 286, in export
    return _export(
        mod,
    ...<5 lines>...
        pre_dispatch=True,
    )

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 1164, in wrapper
    raise e

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 1130, in wrapper
    ep = fn(*args, **kwargs)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/exported_program.py", line 123, in wrapper
    return fn(*args, **kwargs)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 2176, in _export
    ep = _export_for_training(
        mod,
    ...<4 lines>...
        preserve_module_call_signature=preserve_module_call_signature,
    )

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 1164, in wrapper
    raise e

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 1130, in wrapper
    ep = fn(*args, **kwargs)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/exported_program.py", line 123, in wrapper
    return fn(*args, **kwargs)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 2037, in _export_for_training
    export_artifact = export_func(
        mod=mod,
    ...<7 lines>...
        _to_aten_func=_export_to_aten_ir_make_fx,
    )

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 1932, in _non_strict_export
    ) = make_fake_inputs(
        ~~~~~~~~~~~~~~~~^
        mod,
        ^^^^
    ...<4 lines>...
        allow_complex_guards_as_runtime_asserts=allow_complex_guards_as_runtime_asserts,  # for shape env initialization
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    )
    ^

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_export/non_strict_utils.py", line 444, in make_fake_inputs
    equalities_inputs = EqualityConstraint(
        source_pairs=source_pairs,
    ...<3 lines>...
        warn_only=False,
    )

  File "<string>", line 8, in __init__

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/fx/experimental/symbolic_shapes.py", line 1935, in __post_init__
    self._defs[self._find(source)] = fn(self._rewrite(root))
                                        ~~~~~~~~~~~~~^^^^^^

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/fx/experimental/symbolic_shapes.py", line 1959, in _rewrite
    return sympy.Symbol(src.name())
                        ^^^^^^^^

AttributeError: 'One' object has no attribute 'name'


# ⚠️ Errors from strategy 'TorchExportStrictStrategy': -----------------------

Traceback (most recent call last):

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/onnx/_internal/exporter/_capture_strategies.py", line 118, in __call__
    exported_program = self._capture(model, args, kwargs, dynamic_shapes)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/onnx/_internal/exporter/_capture_strategies.py", line 177, in _capture
    return torch.export.export(
           ~~~~~~~~~~~~~~~~~~~^
        model, args, kwargs=kwargs, dynamic_shapes=new_shapes, strict=True
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    )
    ^

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/__init__.py", line 319, in export
    raise e

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/__init__.py", line 286, in export
    return _export(
        mod,
    ...<5 lines>...
        pre_dispatch=True,
    )

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 1164, in wrapper
    raise e

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 1130, in wrapper
    ep = fn(*args, **kwargs)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/exported_program.py", line 123, in wrapper
    return fn(*args, **kwargs)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 2176, in _export
    ep = _export_for_training(
        mod,
    ...<4 lines>...
        preserve_module_call_signature=preserve_module_call_signature,
    )

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 1164, in wrapper
    raise e

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 1130, in wrapper
    ep = fn(*args, **kwargs)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/exported_program.py", line 123, in wrapper
    return fn(*args, **kwargs)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 2037, in _export_for_training
    export_artifact = export_func(
        mod=mod,
    ...<7 lines>...
        _to_aten_func=_export_to_aten_ir_make_fx,
    )

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 1411, in _strict_export
    gm_torch_level = _export_to_torch_ir(
        mod,
    ...<6 lines>...
        _log_export_usage=False,
    )

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/export/_trace.py", line 797, in _export_to_torch_ir
    gm_torch_level, _ = torch._dynamo.export(
                        ~~~~~~~~~~~~~~~~~~~~~
    ...<11 lines>...
        same_signature=same_signature,
        ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    )(
    ~^
        *args,
        ^^^^^^
        **kwargs,
        ^^^^^^^^^
    )
    ^

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 1836, in inner
    result_traced = opt_f(*args, **kwargs)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 375, in __call__
    return super().__call__(*args, **kwargs)
           ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1773, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1784, in _call_impl
    return forward_call(*args, **kwargs)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_dynamo/eval_frame.py", line 736, in compile_wrapper
    return fn(*args, **kwargs)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1773, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py", line 1784, in _call_impl
    return forward_call(*args, **kwargs)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 1495, in __call__
    return self._torchdynamo_orig_callable(
           ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^
        frame, cache_entry, self.hooks, frame_state, skip=1
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    )
    ^

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 629, in __call__
    return _compile(
        frame.f_code,
    ...<15 lines>...
        package=self._package,
    )

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 1164, in _compile
    raise InternalTorchDynamoError(
        f"{type(e).__qualname__}: {str(e)}"
    ).with_traceback(e.__traceback__) from None

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 1111, in _compile
    guarded_code = compile_inner(code, one_graph, hooks, transform)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_utils_internal.py", line 97, in wrapper_function
    return function(*args, **kwargs)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 793, in compile_inner
    return _compile_inner(code, one_graph, hooks, transform)

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 942, in _compile_inner
    check_fn = CheckFunctionManager(
        code,
    ...<4 lines>...
        guards_serialization_mode="save" if package else None,
    )

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_dynamo/guards.py", line 2798, in __init__
    builder, guard_manager = self.build_guards(
                             ~~~~~~~~~~~~~~~~~^
        sorted_guards,
        ^^^^^^^^^^^^^^
    ...<3 lines>...
        None if guard_filter_fn else self.guards_serialization_mode,
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    )
    ^

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_dynamo/guards.py", line 3056, in build_guards
    guard.create(builder)
    ~~~~~~~~~~~~^^^^^^^^^

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_guards.py", line 360, in create
    return self.create_fn(builder, self)
           ~~~~~~~~~~~~~~^^^^^^^^^^^^^^^

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_dynamo/guards.py", line 2073, in SHAPE_ENV
    equalities_inputs = EqualityConstraint(
        source_pairs=source_pairs,
    ...<3 lines>...
        warn_only=False,
    )

  File "<string>", line 8, in __init__

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/fx/experimental/symbolic_shapes.py", line 1935, in __post_init__
    self._defs[self._find(source)] = fn(self._rewrite(root))
                                        ~~~~~~~~~~~~~^^^^^^

  File "/home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/fx/experimental/symbolic_shapes.py", line 1959, in _rewrite
    return sympy.Symbol(src.name())
                        ^^^^^^^^

torch._dynamo.exc.InternalTorchDynamoError: AttributeError: 'One' object has no attribute 'name'



# ⚠️ Errors from strategy 'TorchExportDraftExportStrategy': -----------------------

RuntimeError: �[93m
###################################################################################################
WARNING: 20 issue(s) found during export, and it was not able to soundly produce a graph.
Please follow the instructions to fix the errors.
###################################################################################################

1. Guard Added.
    A guard was added during tracing, which might've resulted in some incorrect
    tracing or constraint violation error.
    Specifically, this guard was added: Eq(Mod(s95 + (Mod(64 - (Mod(s95, 64)), 64)) - 2, 2), 0), where {'s95': "L['I'].size()[2]"}.
    This occurred at the following stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1773, in _wrapped_call_impl
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1784, in _call_impl
            x_128 = self.pool(extr):
        
        Locals:


    And the following framework stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 4578, in pooling_output_shape
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 4551, in pooling_output_shape_pad_lr
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 4536, in div_rtn
            if r != 0 and (bool(r < 0) != bool(y < 0)):

    Because of this, we have modified the dynamic shapes structure to be the
    following. You can also use torch.export.Dim.AUTO instead to specify your
    dynamic shapes, and we will automatically infer the dynamism for you.
    ```
    dynamic_shapes = {'I': {2: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False), 3: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False)}}
    ```

2. Guard Added.
    A guard was added during tracing, which might've resulted in some incorrect
    tracing or constraint violation error.
    Specifically, this guard was added: Eq(Mod((((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//2)) - 2, 2), 0), where {'s95': "L['I'].size()[2]"}.
    This occurred at the following stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1773, in _wrapped_call_impl
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1784, in _call_impl
            x_64  = self.pool(F.relu(self.enc1(x_128))):
        
        Locals:


    And the following framework stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 4578, in pooling_output_shape
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 4551, in pooling_output_shape_pad_lr
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 4536, in div_rtn
            if r != 0 and (bool(r < 0) != bool(y < 0)):

    Because of this, we have modified the dynamic shapes structure to be the
    following. You can also use torch.export.Dim.AUTO instead to specify your
    dynamic shapes, and we will automatically infer the dynamism for you.
    ```
    dynamic_shapes = {'I': {2: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False), 3: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False)}}
    ```

3. Guard Added.
    A guard was added during tracing, which might've resulted in some incorrect
    tracing or constraint violation error.
    Specifically, this guard was added: Eq(Mod((((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//4)) - 2, 2), 0), where {'s95': "L['I'].size()[2]"}.
    This occurred at the following stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1773, in _wrapped_call_impl
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1784, in _call_impl
            x_32  = self.pool(F.relu(self.enc2(x_64))):
        
        Locals:


    And the following framework stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 4578, in pooling_output_shape
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 4551, in pooling_output_shape_pad_lr
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 4536, in div_rtn
            if r != 0 and (bool(r < 0) != bool(y < 0)):

    Because of this, we have modified the dynamic shapes structure to be the
    following. You can also use torch.export.Dim.AUTO instead to specify your
    dynamic shapes, and we will automatically infer the dynamism for you.
    ```
    dynamic_shapes = {'I': {2: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False), 3: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False)}}
    ```

4. Guard Added.
    A guard was added during tracing, which might've resulted in some incorrect
    tracing or constraint violation error.
    Specifically, this guard was added: Eq(Mod((((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//8)) - 2, 2), 0), where {'s95': "L['I'].size()[2]"}.
    This occurred at the following stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1773, in _wrapped_call_impl
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1784, in _call_impl
            x_16  = self.pool(F.relu(self.enc3(x_32))):
        
        Locals:


    And the following framework stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 4578, in pooling_output_shape
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 4551, in pooling_output_shape_pad_lr
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 4536, in div_rtn
            if r != 0 and (bool(r < 0) != bool(y < 0)):

    Because of this, we have modified the dynamic shapes structure to be the
    following. You can also use torch.export.Dim.AUTO instead to specify your
    dynamic shapes, and we will automatically infer the dynamism for you.
    ```
    dynamic_shapes = {'I': {2: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False), 3: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False)}}
    ```

5. Guard Added.
    A guard was added during tracing, which might've resulted in some incorrect
    tracing or constraint violation error.
    Specifically, this guard was added: Eq(Mod((((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//16)) - 2, 2), 0), where {'s95': "L['I'].size()[2]"}.
    This occurred at the following stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1773, in _wrapped_call_impl
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1784, in _call_impl
            x_8   = self.pool(F.relu(self.enc4(x_16))):
        
        Locals:


    And the following framework stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 4578, in pooling_output_shape
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 4551, in pooling_output_shape_pad_lr
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 4536, in div_rtn
            if r != 0 and (bool(r < 0) != bool(y < 0)):

    Because of this, we have modified the dynamic shapes structure to be the
    following. You can also use torch.export.Dim.AUTO instead to specify your
    dynamic shapes, and we will automatically infer the dynamism for you.
    ```
    dynamic_shapes = {'I': {2: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False), 3: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False)}}
    ```

6. Guard Added.
    A guard was added during tracing, which might've resulted in some incorrect
    tracing or constraint violation error.
    Specifically, this guard was added: Eq(Mod((((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//32)) - 2, 2), 0), where {'s95': "L['I'].size()[2]"}.
    This occurred at the following stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1773, in _wrapped_call_impl
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1784, in _call_impl
            x_4   = self.pool(F.relu(self.enc5(x_8))):
        
        Locals:


    And the following framework stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 4578, in pooling_output_shape
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 4551, in pooling_output_shape_pad_lr
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 4536, in div_rtn
            if r != 0 and (bool(r < 0) != bool(y < 0)):

    Because of this, we have modified the dynamic shapes structure to be the
    following. You can also use torch.export.Dim.AUTO instead to specify your
    dynamic shapes, and we will automatically infer the dynamism for you.
    ```
    dynamic_shapes = {'I': {2: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False), 3: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False)}}
    ```

7. Guard Added.
    A guard was added during tracing, which might've resulted in some incorrect
    tracing or constraint violation error.
    Specifically, this guard was added: Max(1, ((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))*Max(1, ((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)) <= 1, where {'s95': "L['I'].size()[2]", 's28': "L['I'].size()[3]"}.
    This occurred at the following stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1773, in _wrapped_call_impl
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1784, in _call_impl
            x     = F.relu(self.dec0(torch.cat([self.upsample(x_4), x_8],   1))):
        
        Locals:


    And the following framework stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_decomp/decompositions.py, lineno 3027, in _upsample_nearest
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_prims_common/__init__.py, lineno 1967, in suggest_memory_format
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_prims_common/__init__.py, lineno 1958, in are_strides_like_channels_last
            if guard_size_oblivious(strides[d] > 1):

    Because of this, we have modified the dynamic shapes structure to be the
    following. You can also use torch.export.Dim.AUTO instead to specify your
    dynamic shapes, and we will automatically infer the dynamism for you.
    ```
    dynamic_shapes = {'I': {2: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False), 3: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False)}}
    ```

8. Guard Added.
    A guard was added during tracing, which might've resulted in some incorrect
    tracing or constraint violation error.
    Specifically, this guard was added: Eq(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//16), 4*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64))), where {'s95': "L['I'].size()[2]"}.
    This occurred at the following stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1773, in _wrapped_call_impl
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1784, in _call_impl
            x     = F.relu(self.dec1(torch.cat([self.upsample(x),   x_16],  1))):
        
        Locals:


    And the following framework stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_ops.py, lineno 829, in __call__
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/utils/_stats.py, lineno 28, in wrapper
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_prims/__init__.py, lineno 1796, in _cat_meta
            torch._check(

    Because of this, we have modified the dynamic shapes structure to be the
    following. You can also use torch.export.Dim.AUTO instead to specify your
    dynamic shapes, and we will automatically infer the dynamism for you.
    ```
    dynamic_shapes = {'I': {2: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False), 3: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False)}}
    ```

9. Guard Added.
    A guard was added during tracing, which might've resulted in some incorrect
    tracing or constraint violation error.
    Specifically, this guard was added: Eq(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//8), 8*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64))), where {'s95': "L['I'].size()[2]"}.
    This occurred at the following stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1773, in _wrapped_call_impl
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1784, in _call_impl
            x     = F.relu(self.dec2(torch.cat([self.upsample(x),   x_32],  1))):
        
        Locals:


    And the following framework stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_ops.py, lineno 829, in __call__
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/utils/_stats.py, lineno 28, in wrapper
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_prims/__init__.py, lineno 1796, in _cat_meta
            torch._check(

    Because of this, we have modified the dynamic shapes structure to be the
    following. You can also use torch.export.Dim.AUTO instead to specify your
    dynamic shapes, and we will automatically infer the dynamism for you.
    ```
    dynamic_shapes = {'I': {2: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False), 3: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False)}}
    ```

10. Guard Added.
    A guard was added during tracing, which might've resulted in some incorrect
    tracing or constraint violation error.
    Specifically, this guard was added: Eq(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//4), 16*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64))), where {'s95': "L['I'].size()[2]"}.
    This occurred at the following stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1773, in _wrapped_call_impl
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1784, in _call_impl
            x     = F.relu(self.dec3(torch.cat([self.upsample(x),   x_64],  1))):
        
        Locals:


    And the following framework stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_ops.py, lineno 829, in __call__
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/utils/_stats.py, lineno 28, in wrapper
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_prims/__init__.py, lineno 1796, in _cat_meta
            torch._check(

    Because of this, we have modified the dynamic shapes structure to be the
    following. You can also use torch.export.Dim.AUTO instead to specify your
    dynamic shapes, and we will automatically infer the dynamism for you.
    ```
    dynamic_shapes = {'I': {2: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False), 3: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False)}}
    ```

11. Guard Added.
    A guard was added during tracing, which might've resulted in some incorrect
    tracing or constraint violation error.
    Specifically, this guard was added: Eq(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//2), 32*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64))), where {'s95': "L['I'].size()[2]"}.
    This occurred at the following stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1773, in _wrapped_call_impl
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1784, in _call_impl
            x     = F.relu(self.dec4(torch.cat([self.upsample(x),   x_128], 1))):
        
        Locals:


    And the following framework stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_ops.py, lineno 829, in __call__
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/utils/_stats.py, lineno 28, in wrapper
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_prims/__init__.py, lineno 1796, in _cat_meta
            torch._check(

    Because of this, we have modified the dynamic shapes structure to be the
    following. You can also use torch.export.Dim.AUTO instead to specify your
    dynamic shapes, and we will automatically infer the dynamism for you.
    ```
    dynamic_shapes = {'I': {2: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False), 3: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False)}}
    ```

12. Guard Added.
    A guard was added during tracing, which might've resulted in some incorrect
    tracing or constraint violation error.
    Specifically, this guard was added: Eq(s95 + (Mod(64 - (Mod(s95, 64)), 64)), 64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64))), where {'s95': "L['I'].size()[2]"}.
    This occurred at the following stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1773, in _wrapped_call_impl
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1784, in _call_impl
            x     = F.relu(self.dec5(torch.cat([self.upsample(x),   I_aligned],  1))):
        
        Locals:


    And the following framework stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_ops.py, lineno 829, in __call__
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/utils/_stats.py, lineno 28, in wrapper
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_prims/__init__.py, lineno 1796, in _cat_meta
            torch._check(

    Because of this, we have modified the dynamic shapes structure to be the
    following. You can also use torch.export.Dim.AUTO instead to specify your
    dynamic shapes, and we will automatically infer the dynamism for you.
    ```
    dynamic_shapes = {'I': {2: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False), 3: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False)}}
    ```

13. Guard Added.
    A guard was added during tracing, which might've resulted in some incorrect
    tracing or constraint violation error.
    Specifically, this guard was added: s95 <= 64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), where {'s95': "L['I'].size()[2]"}.
    This occurred at the following stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1773, in _wrapped_call_impl
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1784, in _call_impl
            x = x[:, :, :H, :W]:
        
        Locals:


    And the following framework stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_ops.py, lineno 829, in __call__
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/utils/_stats.py, lineno 28, in wrapper
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_decomp/decompositions.py, lineno 757, in slice_forward
            elif guard_size_oblivious(end_val > sizes[dim]):

    Because of this, we have modified the dynamic shapes structure to be the
    following. You can also use torch.export.Dim.AUTO instead to specify your
    dynamic shapes, and we will automatically infer the dynamism for you.
    ```
    dynamic_shapes = {'I': {2: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False), 3: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False)}}
    ```

14. Guard Added.
    A guard was added during tracing, which might've resulted in some incorrect
    tracing or constraint violation error.
    Specifically, this guard was added: (s28 + (Mod(64 - (Mod(s28, 64)), 64)))*(3*s95 + 3*(Mod(64 - (Mod(s95, 64)), 64))) <= 2147483647, where {'s95': "L['I'].size()[2]", 's28': "L['I'].size()[3]"}.
    This occurred at the following stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1773, in _wrapped_call_impl
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1784, in _call_impl
            extr = F.relu(self.enc0(I_aligned)):
        
        Locals:


    And the following framework stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/utils/_stats.py, lineno 28, in wrapper
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_subclasses/fake_impls.py, lineno 163, in dispatch_to_op_implementations_dict
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_subclasses/fake_impls.py, lineno 802, in conv
            conv_backend = torch._C._select_conv_backend(**kwargs)

    Because of this, we have modified the dynamic shapes structure to be the
    following. You can also use torch.export.Dim.AUTO instead to specify your
    dynamic shapes, and we will automatically infer the dynamism for you.
    ```
    dynamic_shapes = {'I': {2: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False), 3: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False)}}
    ```

15. Guard Added.
    A guard was added during tracing, which might've resulted in some incorrect
    tracing or constraint violation error.
    Specifically, this guard was added: Eq(Max(1, 2*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))), 2*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))), where {'s28': "L['I'].size()[3]"}.
    This occurred at the following stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1773, in _wrapped_call_impl
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1784, in _call_impl
            x     = F.relu(self.con0(x)):
        
        Locals:


    And the following framework stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_ops.py, lineno 829, in __call__
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 2639, in meta_conv
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 2617, in pick_memory_format
            if input_tensor.is_contiguous(memory_format=torch.contiguous_format):

    Because of this, we have modified the dynamic shapes structure to be the
    following. You can also use torch.export.Dim.AUTO instead to specify your
    dynamic shapes, and we will automatically infer the dynamism for you.
    ```
    dynamic_shapes = {'I': {2: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False), 3: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False)}}
    ```

16. Guard Added.
    A guard was added during tracing, which might've resulted in some incorrect
    tracing or constraint violation error.
    Specifically, this guard was added: Eq(Max(1, 4*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))), 4*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))), where {'s28': "L['I'].size()[3]"}.
    This occurred at the following stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1773, in _wrapped_call_impl
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1784, in _call_impl
            x     = F.relu(self.con1(x)):
        
        Locals:


    And the following framework stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_ops.py, lineno 829, in __call__
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 2639, in meta_conv
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 2617, in pick_memory_format
            if input_tensor.is_contiguous(memory_format=torch.contiguous_format):

    Because of this, we have modified the dynamic shapes structure to be the
    following. You can also use torch.export.Dim.AUTO instead to specify your
    dynamic shapes, and we will automatically infer the dynamism for you.
    ```
    dynamic_shapes = {'I': {2: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False), 3: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False)}}
    ```

17. Guard Added.
    A guard was added during tracing, which might've resulted in some incorrect
    tracing or constraint violation error.
    Specifically, this guard was added: Eq(Max(1, 8*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))), 8*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))), where {'s28': "L['I'].size()[3]"}.
    This occurred at the following stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1773, in _wrapped_call_impl
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1784, in _call_impl
            x     = F.relu(self.con2(x)):
        
        Locals:


    And the following framework stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_ops.py, lineno 829, in __call__
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 2639, in meta_conv
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 2617, in pick_memory_format
            if input_tensor.is_contiguous(memory_format=torch.contiguous_format):

    Because of this, we have modified the dynamic shapes structure to be the
    following. You can also use torch.export.Dim.AUTO instead to specify your
    dynamic shapes, and we will automatically infer the dynamism for you.
    ```
    dynamic_shapes = {'I': {2: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False), 3: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False)}}
    ```

18. Guard Added.
    A guard was added during tracing, which might've resulted in some incorrect
    tracing or constraint violation error.
    Specifically, this guard was added: Eq(Max(1, 16*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))), 16*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))), where {'s28': "L['I'].size()[3]"}.
    This occurred at the following stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1773, in _wrapped_call_impl
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1784, in _call_impl
            x     = F.relu(self.con3(x)):
        
        Locals:


    And the following framework stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_ops.py, lineno 829, in __call__
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 2639, in meta_conv
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 2617, in pick_memory_format
            if input_tensor.is_contiguous(memory_format=torch.contiguous_format):

    Because of this, we have modified the dynamic shapes structure to be the
    following. You can also use torch.export.Dim.AUTO instead to specify your
    dynamic shapes, and we will automatically infer the dynamism for you.
    ```
    dynamic_shapes = {'I': {2: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False), 3: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False)}}
    ```

19. Guard Added.
    A guard was added during tracing, which might've resulted in some incorrect
    tracing or constraint violation error.
    Specifically, this guard was added: Eq(Max(1, 32*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))), 32*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))), where {'s28': "L['I'].size()[3]"}.
    This occurred at the following stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1773, in _wrapped_call_impl
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1784, in _call_impl
            x     = F.relu(self.con4(x)):
        
        Locals:


    And the following framework stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_ops.py, lineno 829, in __call__
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 2639, in meta_conv
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 2617, in pick_memory_format
            if input_tensor.is_contiguous(memory_format=torch.contiguous_format):

    Because of this, we have modified the dynamic shapes structure to be the
    following. You can also use torch.export.Dim.AUTO instead to specify your
    dynamic shapes, and we will automatically infer the dynamism for you.
    ```
    dynamic_shapes = {'I': {2: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False), 3: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False)}}
    ```

20. Guard Added.
    A guard was added during tracing, which might've resulted in some incorrect
    tracing or constraint violation error.
    Specifically, this guard was added: Eq(Max(1, 64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))), 64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))), where {'s28': "L['I'].size()[3]"}.
    This occurred at the following stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1773, in _wrapped_call_impl
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/module.py, lineno 1784, in _call_impl
            x     = F.relu(self.con5(x)):
        
        Locals:


    And the following framework stacktrace: 
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_ops.py, lineno 829, in __call__
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 2639, in meta_conv
        File /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/_meta_registrations.py, lineno 2617, in pick_memory_format
            if input_tensor.is_contiguous(memory_format=torch.contiguous_format):

    Because of this, we have modified the dynamic shapes structure to be the
    following. You can also use torch.export.Dim.AUTO instead to specify your
    dynamic shapes, and we will automatically infer the dynamism for you.
    ```
    dynamic_shapes = {'I': {2: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False), 3: _DimHint(type=<_DimHintType.AUTO: 1>, min=64, max=128, _factory=False)}}
    ```

�[0m

Exported program

ExportedProgram:
    class GraphModule(torch.nn.Module):
        def forward(self, p_enc0_weight: "f32[32, 3, 3, 3]", p_enc0_bias: "f32[32]", p_enc1_weight: "f32[48, 32, 3, 3]", p_enc1_bias: "f32[48]", p_enc2_weight: "f32[64, 48, 3, 3]", p_enc2_bias: "f32[64]", p_enc3_weight: "f32[80, 64, 3, 3]", p_enc3_bias: "f32[80]", p_enc4_weight: "f32[112, 80, 3, 3]", p_enc4_bias: "f32[112]", p_enc5_weight: "f32[112, 112, 3, 3]", p_enc5_bias: "f32[112]", p_dec0_weight: "f32[112, 224, 3, 3]", p_dec0_bias: "f32[112]", p_con0_weight: "f32[112, 112, 3, 3]", p_con0_bias: "f32[112]", p_dec1_weight: "f32[80, 192, 3, 3]", p_dec1_bias: "f32[80]", p_con1_weight: "f32[80, 80, 3, 3]", p_con1_bias: "f32[80]", p_dec2_weight: "f32[64, 144, 3, 3]", p_dec2_bias: "f32[64]", p_con2_weight: "f32[64, 64, 3, 3]", p_con2_bias: "f32[64]", p_dec3_weight: "f32[48, 112, 3, 3]", p_dec3_bias: "f32[48]", p_con3_weight: "f32[48, 48, 3, 3]", p_con3_bias: "f32[48]", p_dec4_weight: "f32[16, 80, 3, 3]", p_dec4_bias: "f32[16]", p_con4_weight: "f32[16, 16, 3, 3]", p_con4_bias: "f32[16]", p_dec5_weight: "f32[12, 19, 3, 3]", p_dec5_bias: "f32[12]", p_con5_weight: "f32[12, 12, 3, 3]", p_con5_bias: "f32[12]", i: "f32[1, 3, s95, s28]"):
             # 
            sym_size_int_66: "Sym(s95)" = torch.ops.aten.sym_size.int(i, 2)
            sym_size_int_67: "Sym(s28)" = torch.ops.aten.sym_size.int(i, 3)
            mod_4: "Sym(Mod(s95, 64))" = sym_size_int_66 % 64
            mul_250: "Sym(-(Mod(s95, 64)))" = -1 * mod_4;  mod_4 = None
            add_240: "Sym(64 - (Mod(s95, 64)))" = 64 + mul_250;  mul_250 = None
            mod_5: "Sym(Mod(64 - (Mod(s95, 64)), 64))" = add_240 % 64;  add_240 = None
            add_241: "Sym(s95 + (Mod(64 - (Mod(s95, 64)), 64)))" = sym_size_int_66 + mod_5
            floordiv_12: "Sym(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//4))" = add_241 // 4
            ge_5: "Sym((((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//4)) >= 1)" = floordiv_12 >= 1;  ge_5 = None
            floordiv_13: "Sym(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//8))" = add_241 // 8
            ge_6: "Sym((((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//8)) >= 1)" = floordiv_13 >= 1;  ge_6 = None
            floordiv_14: "Sym(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//16))" = add_241 // 16
            ge_7: "Sym((((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//16)) >= 1)" = floordiv_14 >= 1;  ge_7 = None
            floordiv_15: "Sym(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//32))" = add_241 // 32
            ge_8: "Sym((((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//32)) >= 1)" = floordiv_15 >= 1;  ge_8 = None
            floordiv_16: "Sym(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64))" = add_241 // 64
            ge_9: "Sym((((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)) >= 1)" = floordiv_16 >= 1;  ge_9 = None
            mul_251: "Sym(2*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)))" = 2 * floordiv_16
            eq_130: "Sym(Eq(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//32), 2*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64))))" = floordiv_15 == mul_251;  mul_251 = eq_130 = None
            mod_6: "Sym(Mod(s28, 64))" = sym_size_int_67 % 64
            mul_252: "Sym(-(Mod(s28, 64)))" = -1 * mod_6;  mod_6 = None
            add_242: "Sym(64 - (Mod(s28, 64)))" = 64 + mul_252;  mul_252 = None
            mod_7: "Sym(Mod(64 - (Mod(s28, 64)), 64))" = add_242 % 64;  add_242 = None
            add_243: "Sym(s28 + (Mod(64 - (Mod(s28, 64)), 64)))" = sym_size_int_67 + mod_7
            floordiv_17: "Sym(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//32))" = add_243 // 32
            floordiv_18: "Sym(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))" = add_243 // 64
            mul_253: "Sym(2*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64)))" = 2 * floordiv_18
            eq_131: "Sym(Eq(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//32), 2*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))))" = floordiv_17 == mul_253;  floordiv_17 = mul_253 = eq_131 = None
            mul_254: "Sym(4*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)))" = 4 * floordiv_16
            eq_132: "Sym(Eq(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//16), 4*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64))))" = floordiv_14 == mul_254;  mul_254 = eq_132 = None
            floordiv_19: "Sym(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//16))" = add_243 // 16
            mul_255: "Sym(4*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64)))" = 4 * floordiv_18
            eq_133: "Sym(Eq(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//16), 4*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))))" = floordiv_19 == mul_255;  floordiv_19 = mul_255 = eq_133 = None
            mul_256: "Sym(8*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)))" = 8 * floordiv_16
            eq_134: "Sym(Eq(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//8), 8*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64))))" = floordiv_13 == mul_256;  mul_256 = eq_134 = None
            floordiv_20: "Sym(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//8))" = add_243 // 8
            mul_257: "Sym(8*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64)))" = 8 * floordiv_18
            eq_135: "Sym(Eq(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//8), 8*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))))" = floordiv_20 == mul_257;  floordiv_20 = mul_257 = eq_135 = None
            mul_258: "Sym(16*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)))" = 16 * floordiv_16
            eq_136: "Sym(Eq(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//4), 16*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64))))" = floordiv_12 == mul_258;  mul_258 = eq_136 = None
            floordiv_21: "Sym(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//4))" = add_243 // 4
            mul_259: "Sym(16*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64)))" = 16 * floordiv_18
            eq_137: "Sym(Eq(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//4), 16*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))))" = floordiv_21 == mul_259;  floordiv_21 = mul_259 = eq_137 = None
            floordiv_22: "Sym(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//2))" = add_241 // 2
            mul_260: "Sym(32*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)))" = 32 * floordiv_16
            eq_138: "Sym(Eq(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//2), 32*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64))))" = floordiv_22 == mul_260;  floordiv_22 = mul_260 = eq_138 = None
            floordiv_23: "Sym(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//2))" = add_243 // 2
            mul_261: "Sym(32*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64)))" = 32 * floordiv_18
            eq_139: "Sym(Eq(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//2), 32*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))))" = floordiv_23 == mul_261;  floordiv_23 = mul_261 = eq_139 = None
            mul_262: "Sym(64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)))" = 64 * floordiv_16
            eq_140: "Sym(Eq(s95 + (Mod(64 - (Mod(s95, 64)), 64)), 64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64))))" = add_241 == mul_262;  add_241 = mul_262 = eq_140 = None
            mul_263: "Sym(64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64)))" = 64 * floordiv_18;  floordiv_18 = None
            eq_141: "Sym(Eq(s28 + (Mod(64 - (Mod(s28, 64)), 64)), 64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))))" = add_243 == mul_263;  add_243 = mul_263 = eq_141 = None
            ge: "Sym((((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//4)) >= 1)" = floordiv_12 >= 1;  floordiv_12 = ge = None
            ge_1: "Sym((((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//8)) >= 1)" = floordiv_13 >= 1;  floordiv_13 = ge_1 = None
            ge_2: "Sym((((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//16)) >= 1)" = floordiv_14 >= 1;  floordiv_14 = ge_2 = None
            ge_3: "Sym((((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//32)) >= 1)" = floordiv_15 >= 1;  floordiv_15 = ge_3 = None
            ge_4: "Sym((((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)) >= 1)" = floordiv_16 >= 1;  floordiv_16 = ge_4 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:42 in forward, code: I_aligned = F.pad(I, (0, pad_w, 0, pad_h), mode="replicate")
            pad: "f32[1, 3, s95 + (Mod(64 - (Mod(s95, 64)), 64)), s28 + (Mod(64 - (Mod(s28, 64)), 64))]" = torch.ops.aten.pad.default(i, [0, mod_7, 0, mod_5], 'replicate');  i = mod_7 = mod_5 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/conv.py:548 in forward, code: return self._conv_forward(input, self.weight, self.bias)
            convolution: "f32[1, 32, s95 + (Mod(64 - (Mod(s95, 64)), 64)), s28 + (Mod(64 - (Mod(s28, 64)), 64))]" = torch.ops.aten.convolution.default(pad, p_enc0_weight, p_enc0_bias, [1, 1], [1, 1], [1, 1], False, [0, 0], 1);  p_enc0_weight = p_enc0_bias = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:44 in forward, code: extr = F.relu(self.enc0(I_aligned))
            relu: "f32[1, 32, s95 + (Mod(64 - (Mod(s95, 64)), 64)), s28 + (Mod(64 - (Mod(s28, 64)), 64))]" = torch.ops.aten.relu.default(convolution);  convolution = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/pooling.py:224 in forward, code: return F.max_pool2d(
            max_pool2d: "f32[1, 32, ((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//2), ((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//2)]" = torch.ops.aten.max_pool2d.default(relu, [2, 2], [2, 2]);  relu = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/conv.py:548 in forward, code: return self._conv_forward(input, self.weight, self.bias)
            convolution_1: "f32[1, 48, ((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//2), ((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//2)]" = torch.ops.aten.convolution.default(max_pool2d, p_enc1_weight, p_enc1_bias, [1, 1], [1, 1], [1, 1], False, [0, 0], 1);  p_enc1_weight = p_enc1_bias = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:46 in forward, code: x_64  = self.pool(F.relu(self.enc1(x_128)))
            relu_1: "f32[1, 48, ((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//2), ((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//2)]" = torch.ops.aten.relu.default(convolution_1);  convolution_1 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/pooling.py:224 in forward, code: return F.max_pool2d(
            max_pool2d_1: "f32[1, 48, ((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//4), ((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//4)]" = torch.ops.aten.max_pool2d.default(relu_1, [2, 2], [2, 2]);  relu_1 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/conv.py:548 in forward, code: return self._conv_forward(input, self.weight, self.bias)
            convolution_2: "f32[1, 64, ((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//4), ((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//4)]" = torch.ops.aten.convolution.default(max_pool2d_1, p_enc2_weight, p_enc2_bias, [1, 1], [1, 1], [1, 1], False, [0, 0], 1);  p_enc2_weight = p_enc2_bias = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:47 in forward, code: x_32  = self.pool(F.relu(self.enc2(x_64)))
            relu_2: "f32[1, 64, ((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//4), ((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//4)]" = torch.ops.aten.relu.default(convolution_2);  convolution_2 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/pooling.py:224 in forward, code: return F.max_pool2d(
            max_pool2d_2: "f32[1, 64, ((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//8), ((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//8)]" = torch.ops.aten.max_pool2d.default(relu_2, [2, 2], [2, 2]);  relu_2 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/conv.py:548 in forward, code: return self._conv_forward(input, self.weight, self.bias)
            convolution_3: "f32[1, 80, ((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//8), ((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//8)]" = torch.ops.aten.convolution.default(max_pool2d_2, p_enc3_weight, p_enc3_bias, [1, 1], [1, 1], [1, 1], False, [0, 0], 1);  p_enc3_weight = p_enc3_bias = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:48 in forward, code: x_16  = self.pool(F.relu(self.enc3(x_32)))
            relu_3: "f32[1, 80, ((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//8), ((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//8)]" = torch.ops.aten.relu.default(convolution_3);  convolution_3 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/pooling.py:224 in forward, code: return F.max_pool2d(
            max_pool2d_3: "f32[1, 80, ((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//16), ((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//16)]" = torch.ops.aten.max_pool2d.default(relu_3, [2, 2], [2, 2]);  relu_3 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/conv.py:548 in forward, code: return self._conv_forward(input, self.weight, self.bias)
            convolution_4: "f32[1, 112, ((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//16), ((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//16)]" = torch.ops.aten.convolution.default(max_pool2d_3, p_enc4_weight, p_enc4_bias, [1, 1], [1, 1], [1, 1], False, [0, 0], 1);  p_enc4_weight = p_enc4_bias = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:49 in forward, code: x_8   = self.pool(F.relu(self.enc4(x_16)))
            relu_4: "f32[1, 112, ((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//16), ((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//16)]" = torch.ops.aten.relu.default(convolution_4);  convolution_4 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/pooling.py:224 in forward, code: return F.max_pool2d(
            max_pool2d_4: "f32[1, 112, ((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//32), ((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//32)]" = torch.ops.aten.max_pool2d.default(relu_4, [2, 2], [2, 2]);  relu_4 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/conv.py:548 in forward, code: return self._conv_forward(input, self.weight, self.bias)
            convolution_5: "f32[1, 112, ((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//32), ((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//32)]" = torch.ops.aten.convolution.default(max_pool2d_4, p_enc5_weight, p_enc5_bias, [1, 1], [1, 1], [1, 1], False, [0, 0], 1);  p_enc5_weight = p_enc5_bias = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:50 in forward, code: x_4   = self.pool(F.relu(self.enc5(x_8)))
            relu_5: "f32[1, 112, ((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//32), ((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//32)]" = torch.ops.aten.relu.default(convolution_5);  convolution_5 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/pooling.py:224 in forward, code: return F.max_pool2d(
            max_pool2d_5: "f32[1, 112, ((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64), ((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64)]" = torch.ops.aten.max_pool2d.default(relu_5, [2, 2], [2, 2]);  relu_5 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/upsampling.py:172 in forward, code: return F.interpolate(
            upsample_nearest2d: "f32[1, 112, 2*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 2*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.upsample_nearest2d.vec(max_pool2d_5, None, [2.0, 2.0]);  max_pool2d_5 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:51 in forward, code: x     = F.relu(self.dec0(torch.cat([self.upsample(x_4), x_8],   1)))
            cat: "f32[1, 224, 2*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 2*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.cat.default([upsample_nearest2d, max_pool2d_4], 1);  upsample_nearest2d = max_pool2d_4 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/conv.py:548 in forward, code: return self._conv_forward(input, self.weight, self.bias)
            convolution_6: "f32[1, 112, 2*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 2*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.convolution.default(cat, p_dec0_weight, p_dec0_bias, [1, 1], [1, 1], [1, 1], False, [0, 0], 1);  cat = p_dec0_weight = p_dec0_bias = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:51 in forward, code: x     = F.relu(self.dec0(torch.cat([self.upsample(x_4), x_8],   1)))
            relu_6: "f32[1, 112, 2*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 2*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.relu.default(convolution_6);  convolution_6 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/conv.py:548 in forward, code: return self._conv_forward(input, self.weight, self.bias)
            convolution_7: "f32[1, 112, 2*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 2*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.convolution.default(relu_6, p_con0_weight, p_con0_bias, [1, 1], [1, 1], [1, 1], False, [0, 0], 1);  relu_6 = p_con0_weight = p_con0_bias = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:52 in forward, code: x     = F.relu(self.con0(x))
            relu_7: "f32[1, 112, 2*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 2*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.relu.default(convolution_7);  convolution_7 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/upsampling.py:172 in forward, code: return F.interpolate(
            upsample_nearest2d_1: "f32[1, 112, 4*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 4*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.upsample_nearest2d.vec(relu_7, None, [2.0, 2.0]);  relu_7 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:53 in forward, code: x     = F.relu(self.dec1(torch.cat([self.upsample(x),   x_16],  1)))
            cat_1: "f32[1, 192, 4*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 4*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.cat.default([upsample_nearest2d_1, max_pool2d_3], 1);  upsample_nearest2d_1 = max_pool2d_3 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/conv.py:548 in forward, code: return self._conv_forward(input, self.weight, self.bias)
            convolution_8: "f32[1, 80, 4*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 4*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.convolution.default(cat_1, p_dec1_weight, p_dec1_bias, [1, 1], [1, 1], [1, 1], False, [0, 0], 1);  cat_1 = p_dec1_weight = p_dec1_bias = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:53 in forward, code: x     = F.relu(self.dec1(torch.cat([self.upsample(x),   x_16],  1)))
            relu_8: "f32[1, 80, 4*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 4*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.relu.default(convolution_8);  convolution_8 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/conv.py:548 in forward, code: return self._conv_forward(input, self.weight, self.bias)
            convolution_9: "f32[1, 80, 4*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 4*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.convolution.default(relu_8, p_con1_weight, p_con1_bias, [1, 1], [1, 1], [1, 1], False, [0, 0], 1);  relu_8 = p_con1_weight = p_con1_bias = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:54 in forward, code: x     = F.relu(self.con1(x))
            relu_9: "f32[1, 80, 4*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 4*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.relu.default(convolution_9);  convolution_9 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/upsampling.py:172 in forward, code: return F.interpolate(
            upsample_nearest2d_2: "f32[1, 80, 8*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 8*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.upsample_nearest2d.vec(relu_9, None, [2.0, 2.0]);  relu_9 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:55 in forward, code: x     = F.relu(self.dec2(torch.cat([self.upsample(x),   x_32],  1)))
            cat_2: "f32[1, 144, 8*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 8*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.cat.default([upsample_nearest2d_2, max_pool2d_2], 1);  upsample_nearest2d_2 = max_pool2d_2 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/conv.py:548 in forward, code: return self._conv_forward(input, self.weight, self.bias)
            convolution_10: "f32[1, 64, 8*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 8*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.convolution.default(cat_2, p_dec2_weight, p_dec2_bias, [1, 1], [1, 1], [1, 1], False, [0, 0], 1);  cat_2 = p_dec2_weight = p_dec2_bias = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:55 in forward, code: x     = F.relu(self.dec2(torch.cat([self.upsample(x),   x_32],  1)))
            relu_10: "f32[1, 64, 8*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 8*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.relu.default(convolution_10);  convolution_10 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/conv.py:548 in forward, code: return self._conv_forward(input, self.weight, self.bias)
            convolution_11: "f32[1, 64, 8*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 8*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.convolution.default(relu_10, p_con2_weight, p_con2_bias, [1, 1], [1, 1], [1, 1], False, [0, 0], 1);  relu_10 = p_con2_weight = p_con2_bias = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:56 in forward, code: x     = F.relu(self.con2(x))
            relu_11: "f32[1, 64, 8*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 8*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.relu.default(convolution_11);  convolution_11 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/upsampling.py:172 in forward, code: return F.interpolate(
            upsample_nearest2d_3: "f32[1, 64, 16*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 16*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.upsample_nearest2d.vec(relu_11, None, [2.0, 2.0]);  relu_11 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:57 in forward, code: x     = F.relu(self.dec3(torch.cat([self.upsample(x),   x_64],  1)))
            cat_3: "f32[1, 112, 16*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 16*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.cat.default([upsample_nearest2d_3, max_pool2d_1], 1);  upsample_nearest2d_3 = max_pool2d_1 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/conv.py:548 in forward, code: return self._conv_forward(input, self.weight, self.bias)
            convolution_12: "f32[1, 48, 16*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 16*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.convolution.default(cat_3, p_dec3_weight, p_dec3_bias, [1, 1], [1, 1], [1, 1], False, [0, 0], 1);  cat_3 = p_dec3_weight = p_dec3_bias = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:57 in forward, code: x     = F.relu(self.dec3(torch.cat([self.upsample(x),   x_64],  1)))
            relu_12: "f32[1, 48, 16*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 16*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.relu.default(convolution_12);  convolution_12 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/conv.py:548 in forward, code: return self._conv_forward(input, self.weight, self.bias)
            convolution_13: "f32[1, 48, 16*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 16*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.convolution.default(relu_12, p_con3_weight, p_con3_bias, [1, 1], [1, 1], [1, 1], False, [0, 0], 1);  relu_12 = p_con3_weight = p_con3_bias = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:58 in forward, code: x     = F.relu(self.con3(x))
            relu_13: "f32[1, 48, 16*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 16*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.relu.default(convolution_13);  convolution_13 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/upsampling.py:172 in forward, code: return F.interpolate(
            upsample_nearest2d_4: "f32[1, 48, 32*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 32*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.upsample_nearest2d.vec(relu_13, None, [2.0, 2.0]);  relu_13 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:59 in forward, code: x     = F.relu(self.dec4(torch.cat([self.upsample(x),   x_128], 1)))
            cat_4: "f32[1, 80, 32*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 32*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.cat.default([upsample_nearest2d_4, max_pool2d], 1);  upsample_nearest2d_4 = max_pool2d = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/conv.py:548 in forward, code: return self._conv_forward(input, self.weight, self.bias)
            convolution_14: "f32[1, 16, 32*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 32*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.convolution.default(cat_4, p_dec4_weight, p_dec4_bias, [1, 1], [1, 1], [1, 1], False, [0, 0], 1);  cat_4 = p_dec4_weight = p_dec4_bias = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:59 in forward, code: x     = F.relu(self.dec4(torch.cat([self.upsample(x),   x_128], 1)))
            relu_14: "f32[1, 16, 32*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 32*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.relu.default(convolution_14);  convolution_14 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/conv.py:548 in forward, code: return self._conv_forward(input, self.weight, self.bias)
            convolution_15: "f32[1, 16, 32*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 32*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.convolution.default(relu_14, p_con4_weight, p_con4_bias, [1, 1], [1, 1], [1, 1], False, [0, 0], 1);  relu_14 = p_con4_weight = p_con4_bias = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:60 in forward, code: x     = F.relu(self.con4(x))
            relu_15: "f32[1, 16, 32*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 32*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.relu.default(convolution_15);  convolution_15 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/upsampling.py:172 in forward, code: return F.interpolate(
            upsample_nearest2d_5: "f32[1, 16, 64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.upsample_nearest2d.vec(relu_15, None, [2.0, 2.0]);  relu_15 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:61 in forward, code: x     = F.relu(self.dec5(torch.cat([self.upsample(x),   I_aligned],  1)))
            cat_5: "f32[1, 19, 64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.cat.default([upsample_nearest2d_5, pad], 1);  upsample_nearest2d_5 = pad = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/conv.py:548 in forward, code: return self._conv_forward(input, self.weight, self.bias)
            convolution_16: "f32[1, 12, 64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.convolution.default(cat_5, p_dec5_weight, p_dec5_bias, [1, 1], [1, 1], [1, 1], False, [0, 0], 1);  cat_5 = p_dec5_weight = p_dec5_bias = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:61 in forward, code: x     = F.relu(self.dec5(torch.cat([self.upsample(x),   I_aligned],  1)))
            relu_16: "f32[1, 12, 64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.relu.default(convolution_16);  convolution_16 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/.venv/lib/python3.13/site-packages/torch/nn/modules/conv.py:548 in forward, code: return self._conv_forward(input, self.weight, self.bias)
            convolution_17: "f32[1, 12, 64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.convolution.default(relu_16, p_con5_weight, p_con5_bias, [1, 1], [1, 1], [1, 1], False, [0, 0], 1);  relu_16 = p_con5_weight = p_con5_bias = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:62 in forward, code: x     = F.relu(self.con5(x))
            relu_17: "f32[1, 12, 64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.relu.default(convolution_17);  convolution_17 = None
            
             # File: /home/kistenklaus/Documents/kit/hiwi/vkcnn/onnx_example.py:64 in forward, code: x = x[:, :, :H, :W]
            slice_1: "f32[1, 12, 64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.slice.Tensor(relu_17);  relu_17 = None
            slice_2: "f32[1, 12, 64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)), 64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.slice.Tensor(slice_1, 1);  slice_1 = None
            slice_3: "f32[1, 12, s95, 64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]" = torch.ops.aten.slice.Tensor(slice_2, 2, None, sym_size_int_66);  slice_2 = sym_size_int_66 = None
            slice_4: "f32[1, 12, s95, s28]" = torch.ops.aten.slice.Tensor(slice_3, 3, None, sym_size_int_67);  slice_3 = sym_size_int_67 = None
            return (slice_4,)
            
Graph signature: 
    # inputs
    p_enc0_weight: PARAMETER target='enc0.weight'
    p_enc0_bias: PARAMETER target='enc0.bias'
    p_enc1_weight: PARAMETER target='enc1.weight'
    p_enc1_bias: PARAMETER target='enc1.bias'
    p_enc2_weight: PARAMETER target='enc2.weight'
    p_enc2_bias: PARAMETER target='enc2.bias'
    p_enc3_weight: PARAMETER target='enc3.weight'
    p_enc3_bias: PARAMETER target='enc3.bias'
    p_enc4_weight: PARAMETER target='enc4.weight'
    p_enc4_bias: PARAMETER target='enc4.bias'
    p_enc5_weight: PARAMETER target='enc5.weight'
    p_enc5_bias: PARAMETER target='enc5.bias'
    p_dec0_weight: PARAMETER target='dec0.weight'
    p_dec0_bias: PARAMETER target='dec0.bias'
    p_con0_weight: PARAMETER target='con0.weight'
    p_con0_bias: PARAMETER target='con0.bias'
    p_dec1_weight: PARAMETER target='dec1.weight'
    p_dec1_bias: PARAMETER target='dec1.bias'
    p_con1_weight: PARAMETER target='con1.weight'
    p_con1_bias: PARAMETER target='con1.bias'
    p_dec2_weight: PARAMETER target='dec2.weight'
    p_dec2_bias: PARAMETER target='dec2.bias'
    p_con2_weight: PARAMETER target='con2.weight'
    p_con2_bias: PARAMETER target='con2.bias'
    p_dec3_weight: PARAMETER target='dec3.weight'
    p_dec3_bias: PARAMETER target='dec3.bias'
    p_con3_weight: PARAMETER target='con3.weight'
    p_con3_bias: PARAMETER target='con3.bias'
    p_dec4_weight: PARAMETER target='dec4.weight'
    p_dec4_bias: PARAMETER target='dec4.bias'
    p_con4_weight: PARAMETER target='con4.weight'
    p_con4_bias: PARAMETER target='con4.bias'
    p_dec5_weight: PARAMETER target='dec5.weight'
    p_dec5_bias: PARAMETER target='dec5.bias'
    p_con5_weight: PARAMETER target='con5.weight'
    p_con5_bias: PARAMETER target='con5.bias'
    i: USER_INPUT
    
    # outputs
    slice_4: USER_OUTPUT
    
Range constraints: {s95: VR[64, 128], s28: VR[64, 128]}

ONNX model

<
    ir_version=10,
    opset_imports={'pkg.onnxscript.torch_lib.common': 1, '': 18},
    producer_name='pytorch',
    producer_version='2.8.0+cu128',
    domain=None,
    model_version=None,
>
graph(
    name=main_graph,
    inputs=(
        %"input"<FLOAT,[1,3,s95,s28]>
    ),
    outputs=(
        %"output"<FLOAT,[1,12,s95,s28]>
    ),
    initializers=(
        %"enc0.weight"<FLOAT,[32,3,3,3]>{TorchTensor(...)},
        %"enc0.bias"<FLOAT,[32]>{TorchTensor(...)},
        %"enc1.weight"<FLOAT,[48,32,3,3]>{TorchTensor(...)},
        %"enc1.bias"<FLOAT,[48]>{TorchTensor(...)},
        %"enc2.weight"<FLOAT,[64,48,3,3]>{TorchTensor(...)},
        %"enc2.bias"<FLOAT,[64]>{TorchTensor(...)},
        %"enc3.weight"<FLOAT,[80,64,3,3]>{TorchTensor(...)},
        %"enc3.bias"<FLOAT,[80]>{TorchTensor(...)},
        %"enc4.weight"<FLOAT,[112,80,3,3]>{TorchTensor(...)},
        %"enc4.bias"<FLOAT,[112]>{TorchTensor(...)},
        %"enc5.weight"<FLOAT,[112,112,3,3]>{TorchTensor(...)},
        %"enc5.bias"<FLOAT,[112]>{TorchTensor(...)},
        %"dec0.weight"<FLOAT,[112,224,3,3]>{TorchTensor(...)},
        %"dec0.bias"<FLOAT,[112]>{TorchTensor(...)},
        %"con0.weight"<FLOAT,[112,112,3,3]>{TorchTensor(...)},
        %"con0.bias"<FLOAT,[112]>{TorchTensor(...)},
        %"dec1.weight"<FLOAT,[80,192,3,3]>{TorchTensor(...)},
        %"dec1.bias"<FLOAT,[80]>{TorchTensor(...)},
        %"con1.weight"<FLOAT,[80,80,3,3]>{TorchTensor(...)},
        %"con1.bias"<FLOAT,[80]>{TorchTensor(...)},
        %"dec2.weight"<FLOAT,[64,144,3,3]>{TorchTensor(...)},
        %"dec2.bias"<FLOAT,[64]>{TorchTensor(...)},
        %"con2.weight"<FLOAT,[64,64,3,3]>{TorchTensor(...)},
        %"con2.bias"<FLOAT,[64]>{TorchTensor(...)},
        %"dec3.weight"<FLOAT,[48,112,3,3]>{TorchTensor(...)},
        %"dec3.bias"<FLOAT,[48]>{TorchTensor(...)},
        %"con3.weight"<FLOAT,[48,48,3,3]>{TorchTensor(...)},
        %"con3.bias"<FLOAT,[48]>{TorchTensor(...)},
        %"dec4.weight"<FLOAT,[16,80,3,3]>{TorchTensor(...)},
        %"dec4.bias"<FLOAT,[16]>{TorchTensor(...)},
        %"con4.weight"<FLOAT,[16,16,3,3]>{TorchTensor(...)},
        %"con4.bias"<FLOAT,[16]>{TorchTensor(...)},
        %"dec5.weight"<FLOAT,[12,19,3,3]>{TorchTensor(...)},
        %"dec5.bias"<FLOAT,[12]>{TorchTensor(...)},
        %"con5.weight"<FLOAT,[12,12,3,3]>{TorchTensor(...)},
        %"con5.bias"<FLOAT,[12]>{TorchTensor(...)}
    ),
) {
      0 |  # node_Shape_0
           %"val_0"<?,?> ⬅️ ::Shape(%"input") {end=3, start=2}
      1 |  # node_sym_size_int_66
           %"sym_size_int_66"<INT64,[]> ⬅️ ::Squeeze(%"val_0")
      2 |  # node_Shape_1
           %"val_1"<?,?> ⬅️ ::Shape(%"input") {end=4, start=3}
      3 |  # node_sym_size_int_67
           %"sym_size_int_67"<INT64,[]> ⬅️ ::Squeeze(%"val_1")
      4 |  # node_Constant_2
           %"val_2"<?,?> ⬅️ ::Constant() {value=Tensor<INT64,[]>(array(64), name=None)}
      5 |  # node_mod_4
           %"mod_4"<INT64,[]> ⬅️ ::Mod(%"sym_size_int_66", %"val_2") {fmod=0}
      6 |  # node_Constant_3
           %"val_3"<?,?> ⬅️ ::Constant() {value=Tensor<INT64,[]>(array(-1), name=None)}
      7 |  # node_mul_250
           %"mul_250"<INT64,[]> ⬅️ ::Mul(%"val_3", %"mod_4")
      8 |  # node_add_240
           %"add_240"<INT64,[]> ⬅️ ::Add(%"val_2", %"mul_250")
      9 |  # node_mod_5
           %"mod_5"<INT64,[]> ⬅️ ::Mod(%"add_240", %"val_2") {fmod=0}
     10 |  # node_add_241
           %"add_241"<INT64,[]> ⬅️ ::Add(%"sym_size_int_66", %"mod_5")
     11 |  # node_Constant_4
           %"val_4"<?,?> ⬅️ ::Constant() {value=Tensor<INT64,[]>(array(4), name=None)}
     12 |  # node_floordiv_12
           %"floordiv_12"<INT64,[]> ⬅️ ::Div(%"add_241", %"val_4")
     13 |  # node_Constant_5
           %"val_5"<?,?> ⬅️ ::Constant() {value=Tensor<INT64,[]>(array(1), name=None)}
     14 |  # node_ge_5
           %"ge_5"<BOOL,[]> ⬅️ ::GreaterOrEqual(%"floordiv_12", %"val_5")
     15 |  # node_Constant_6
           %"val_6"<?,?> ⬅️ ::Constant() {value=Tensor<INT64,[]>(array(8), name=None)}
     16 |  # node_floordiv_13
           %"floordiv_13"<INT64,[]> ⬅️ ::Div(%"add_241", %"val_6")
     17 |  # node_ge_6
           %"ge_6"<BOOL,[]> ⬅️ ::GreaterOrEqual(%"floordiv_13", %"val_5")
     18 |  # node_Constant_7
           %"val_7"<?,?> ⬅️ ::Constant() {value=Tensor<INT64,[]>(array(16), name=None)}
     19 |  # node_floordiv_14
           %"floordiv_14"<INT64,[]> ⬅️ ::Div(%"add_241", %"val_7")
     20 |  # node_ge_7
           %"ge_7"<BOOL,[]> ⬅️ ::GreaterOrEqual(%"floordiv_14", %"val_5")
     21 |  # node_Constant_8
           %"val_8"<?,?> ⬅️ ::Constant() {value=Tensor<INT64,[]>(array(32), name=None)}
     22 |  # node_floordiv_15
           %"floordiv_15"<INT64,[]> ⬅️ ::Div(%"add_241", %"val_8")
     23 |  # node_ge_8
           %"ge_8"<BOOL,[]> ⬅️ ::GreaterOrEqual(%"floordiv_15", %"val_5")
     24 |  # node_floordiv_16
           %"floordiv_16"<INT64,[]> ⬅️ ::Div(%"add_241", %"val_2")
     25 |  # node_ge_9
           %"ge_9"<BOOL,[]> ⬅️ ::GreaterOrEqual(%"floordiv_16", %"val_5")
     26 |  # node_Constant_9
           %"val_9"<?,?> ⬅️ ::Constant() {value=Tensor<INT64,[]>(array(2), name=None)}
     27 |  # node_mul_251
           %"mul_251"<INT64,[]> ⬅️ ::Mul(%"val_9", %"floordiv_16")
     28 |  # node_eq_130
           %"eq_130"<BOOL,[]> ⬅️ ::Equal(%"floordiv_15", %"mul_251")
     29 |  # node_mod_6
           %"mod_6"<INT64,[]> ⬅️ ::Mod(%"sym_size_int_67", %"val_2") {fmod=0}
     30 |  # node_mul_252
           %"mul_252"<INT64,[]> ⬅️ ::Mul(%"val_3", %"mod_6")
     31 |  # node_add_242
           %"add_242"<INT64,[]> ⬅️ ::Add(%"val_2", %"mul_252")
     32 |  # node_mod_7
           %"mod_7"<INT64,[]> ⬅️ ::Mod(%"add_242", %"val_2") {fmod=0}
     33 |  # node_add_243
           %"add_243"<INT64,[]> ⬅️ ::Add(%"sym_size_int_67", %"mod_7")
     34 |  # node_floordiv_17
           %"floordiv_17"<INT64,[]> ⬅️ ::Div(%"add_243", %"val_8")
     35 |  # node_floordiv_18
           %"floordiv_18"<INT64,[]> ⬅️ ::Div(%"add_243", %"val_2")
     36 |  # node_mul_253
           %"mul_253"<INT64,[]> ⬅️ ::Mul(%"val_9", %"floordiv_18")
     37 |  # node_eq_131
           %"eq_131"<BOOL,[]> ⬅️ ::Equal(%"floordiv_17", %"mul_253")
     38 |  # node_mul_254
           %"mul_254"<INT64,[]> ⬅️ ::Mul(%"val_4", %"floordiv_16")
     39 |  # node_eq_132
           %"eq_132"<BOOL,[]> ⬅️ ::Equal(%"floordiv_14", %"mul_254")
     40 |  # node_floordiv_19
           %"floordiv_19"<INT64,[]> ⬅️ ::Div(%"add_243", %"val_7")
     41 |  # node_mul_255
           %"mul_255"<INT64,[]> ⬅️ ::Mul(%"val_4", %"floordiv_18")
     42 |  # node_eq_133
           %"eq_133"<BOOL,[]> ⬅️ ::Equal(%"floordiv_19", %"mul_255")
     43 |  # node_mul_256
           %"mul_256"<INT64,[]> ⬅️ ::Mul(%"val_6", %"floordiv_16")
     44 |  # node_eq_134
           %"eq_134"<BOOL,[]> ⬅️ ::Equal(%"floordiv_13", %"mul_256")
     45 |  # node_floordiv_20
           %"floordiv_20"<INT64,[]> ⬅️ ::Div(%"add_243", %"val_6")
     46 |  # node_mul_257
           %"mul_257"<INT64,[]> ⬅️ ::Mul(%"val_6", %"floordiv_18")
     47 |  # node_eq_135
           %"eq_135"<BOOL,[]> ⬅️ ::Equal(%"floordiv_20", %"mul_257")
     48 |  # node_mul_258
           %"mul_258"<INT64,[]> ⬅️ ::Mul(%"val_7", %"floordiv_16")
     49 |  # node_eq_136
           %"eq_136"<BOOL,[]> ⬅️ ::Equal(%"floordiv_12", %"mul_258")
     50 |  # node_floordiv_21
           %"floordiv_21"<INT64,[]> ⬅️ ::Div(%"add_243", %"val_4")
     51 |  # node_mul_259
           %"mul_259"<INT64,[]> ⬅️ ::Mul(%"val_7", %"floordiv_18")
     52 |  # node_eq_137
           %"eq_137"<BOOL,[]> ⬅️ ::Equal(%"floordiv_21", %"mul_259")
     53 |  # node_floordiv_22
           %"floordiv_22"<INT64,[]> ⬅️ ::Div(%"add_241", %"val_9")
     54 |  # node_mul_260
           %"mul_260"<INT64,[]> ⬅️ ::Mul(%"val_8", %"floordiv_16")
     55 |  # node_eq_138
           %"eq_138"<BOOL,[]> ⬅️ ::Equal(%"floordiv_22", %"mul_260")
     56 |  # node_floordiv_23
           %"floordiv_23"<INT64,[]> ⬅️ ::Div(%"add_243", %"val_9")
     57 |  # node_mul_261
           %"mul_261"<INT64,[]> ⬅️ ::Mul(%"val_8", %"floordiv_18")
     58 |  # node_eq_139
           %"eq_139"<BOOL,[]> ⬅️ ::Equal(%"floordiv_23", %"mul_261")
     59 |  # node_mul_262
           %"mul_262"<INT64,[]> ⬅️ ::Mul(%"val_2", %"floordiv_16")
     60 |  # node_eq_140
           %"eq_140"<BOOL,[]> ⬅️ ::Equal(%"add_241", %"mul_262")
     61 |  # node_mul_263
           %"mul_263"<INT64,[]> ⬅️ ::Mul(%"val_2", %"floordiv_18")
     62 |  # node_eq_141
           %"eq_141"<BOOL,[]> ⬅️ ::Equal(%"add_243", %"mul_263")
     63 |  # node_ge
           %"ge"<BOOL,[]> ⬅️ ::GreaterOrEqual(%"floordiv_12", %"val_5")
     64 |  # node_ge_1
           %"ge_1"<BOOL,[]> ⬅️ ::GreaterOrEqual(%"floordiv_13", %"val_5")
     65 |  # node_ge_2
           %"ge_2"<BOOL,[]> ⬅️ ::GreaterOrEqual(%"floordiv_14", %"val_5")
     66 |  # node_ge_3
           %"ge_3"<BOOL,[]> ⬅️ ::GreaterOrEqual(%"floordiv_15", %"val_5")
     67 |  # node_ge_4
           %"ge_4"<BOOL,[]> ⬅️ ::GreaterOrEqual(%"floordiv_16", %"val_5")
     68 |  # node_Constant_10
           %"val_10"<?,?> ⬅️ ::Constant() {value_ints=[0]}
     69 |  # node_Constant_11
           %"val_11"<?,?> ⬅️ ::Constant() {value=Tensor<INT64,[1]>(array([-1]), name=None)}
     70 |  # node_Reshape_12
           %"val_12"<?,?> ⬅️ ::Reshape(%"mod_7", %"val_11") {allowzero=0}
     71 |  # node_Constant_13
           %"val_13"<?,?> ⬅️ ::Constant() {value_ints=[0]}
     72 |  # node_Reshape_14
           %"val_14"<?,?> ⬅️ ::Reshape(%"mod_5", %"val_11") {allowzero=0}
     73 |  # node_Constant_15
           %"val_15"<?,?> ⬅️ ::Constant() {value_ints=[0]}
     74 |  # node_Concat_16
           %"val_16"<?,?> ⬅️ ::Concat(%"val_15", %"val_15", %"val_13", %"val_10", %"val_15", %"val_15", %"val_14", %"val_12") {axis=0}
     75 |  # node_pad
           %"pad"<FLOAT,[1,3,s95 + (Mod(64 - (Mod(s95, 64)), 64)),s28 + (Mod(64 - (Mod(s28, 64)), 64))]> ⬅️ ::Pad(%"input", %"val_16") {mode=edge}
     76 |  # node_convolution
           %"convolution"<FLOAT,[1,32,s95 + (Mod(64 - (Mod(s95, 64)), 64)),s28 + (Mod(64 - (Mod(s28, 64)), 64))]> ⬅️ ::Conv(%"pad", %"enc0.weight"{...}, %"enc0.bias"{...}) {group=1, pads=[1, 1, 1, 1], auto_pad=NOTSET, strides=[1, 1], dilations=[1, 1]}
     77 |  # node_relu
           %"relu"<FLOAT,[1,32,s95 + (Mod(64 - (Mod(s95, 64)), 64)),s28 + (Mod(64 - (Mod(s28, 64)), 64))]> ⬅️ ::Relu(%"convolution")
     78 |  # node_max_pool2d
           %"max_pool2d"<FLOAT,[1,32,((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//2),((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//2)]>, %"val_17"<?,?> ⬅️ ::MaxPool(%"relu") {storage_order=0, dilations=(1, 1), ceil_mode=False, pads=(0, 0, 0, 0), auto_pad=NOTSET, strides=[2, 2], kernel_shape=[2, 2]}
     79 |  # node_convolution_1
           %"convolution_1"<FLOAT,[1,48,((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//2),((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//2)]> ⬅️ ::Conv(%"max_pool2d", %"enc1.weight"{...}, %"enc1.bias"{...}) {group=1, pads=[1, 1, 1, 1], auto_pad=NOTSET, strides=[1, 1], dilations=[1, 1]}
     80 |  # node_relu_1
           %"relu_1"<FLOAT,[1,48,((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//2),((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//2)]> ⬅️ ::Relu(%"convolution_1")
     81 |  # node_max_pool2d_1
           %"max_pool2d_1"<FLOAT,[1,48,((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//4),((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//4)]>, %"val_18"<?,?> ⬅️ ::MaxPool(%"relu_1") {storage_order=0, dilations=(1, 1), ceil_mode=False, pads=(0, 0, 0, 0), auto_pad=NOTSET, strides=[2, 2], kernel_shape=[2, 2]}
     82 |  # node_convolution_2
           %"convolution_2"<FLOAT,[1,64,((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//4),((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//4)]> ⬅️ ::Conv(%"max_pool2d_1", %"enc2.weight"{...}, %"enc2.bias"{...}) {group=1, pads=[1, 1, 1, 1], auto_pad=NOTSET, strides=[1, 1], dilations=[1, 1]}
     83 |  # node_relu_2
           %"relu_2"<FLOAT,[1,64,((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//4),((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//4)]> ⬅️ ::Relu(%"convolution_2")
     84 |  # node_max_pool2d_2
           %"max_pool2d_2"<FLOAT,[1,64,((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//8),((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//8)]>, %"val_19"<?,?> ⬅️ ::MaxPool(%"relu_2") {storage_order=0, dilations=(1, 1), ceil_mode=False, pads=(0, 0, 0, 0), auto_pad=NOTSET, strides=[2, 2], kernel_shape=[2, 2]}
     85 |  # node_convolution_3
           %"convolution_3"<FLOAT,[1,80,((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//8),((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//8)]> ⬅️ ::Conv(%"max_pool2d_2", %"enc3.weight"{...}, %"enc3.bias"{...}) {group=1, pads=[1, 1, 1, 1], auto_pad=NOTSET, strides=[1, 1], dilations=[1, 1]}
     86 |  # node_relu_3
           %"relu_3"<FLOAT,[1,80,((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//8),((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//8)]> ⬅️ ::Relu(%"convolution_3")
     87 |  # node_max_pool2d_3
           %"max_pool2d_3"<FLOAT,[1,80,((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//16),((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//16)]>, %"val_20"<?,?> ⬅️ ::MaxPool(%"relu_3") {storage_order=0, dilations=(1, 1), ceil_mode=False, pads=(0, 0, 0, 0), auto_pad=NOTSET, strides=[2, 2], kernel_shape=[2, 2]}
     88 |  # node_convolution_4
           %"convolution_4"<FLOAT,[1,112,((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//16),((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//16)]> ⬅️ ::Conv(%"max_pool2d_3", %"enc4.weight"{...}, %"enc4.bias"{...}) {group=1, pads=[1, 1, 1, 1], auto_pad=NOTSET, strides=[1, 1], dilations=[1, 1]}
     89 |  # node_relu_4
           %"relu_4"<FLOAT,[1,112,((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//16),((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//16)]> ⬅️ ::Relu(%"convolution_4")
     90 |  # node_max_pool2d_4
           %"max_pool2d_4"<FLOAT,[1,112,((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//32),((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//32)]>, %"val_21"<?,?> ⬅️ ::MaxPool(%"relu_4") {storage_order=0, dilations=(1, 1), ceil_mode=False, pads=(0, 0, 0, 0), auto_pad=NOTSET, strides=[2, 2], kernel_shape=[2, 2]}
     91 |  # node_convolution_5
           %"convolution_5"<FLOAT,[1,112,((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//32),((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//32)]> ⬅️ ::Conv(%"max_pool2d_4", %"enc5.weight"{...}, %"enc5.bias"{...}) {group=1, pads=[1, 1, 1, 1], auto_pad=NOTSET, strides=[1, 1], dilations=[1, 1]}
     92 |  # node_relu_5
           %"relu_5"<FLOAT,[1,112,((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//32),((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//32)]> ⬅️ ::Relu(%"convolution_5")
     93 |  # node_max_pool2d_5
           %"max_pool2d_5"<FLOAT,[1,112,((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64),((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64)]>, %"val_22"<?,?> ⬅️ ::MaxPool(%"relu_5") {storage_order=0, dilations=(1, 1), ceil_mode=False, pads=(0, 0, 0, 0), auto_pad=NOTSET, strides=[2, 2], kernel_shape=[2, 2]}
     94 |  # node_Constant_17
           %"val_23"<?,?> ⬅️ ::Constant() {value_floats=[1.0, 1.0, 2.0, 2.0]}
     95 |  # node_upsample_nearest2d
           %"upsample_nearest2d"<FLOAT,[1,112,2*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),2*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Resize(%"max_pool2d_5", None, %"val_23") {keep_aspect_ratio_policy=stretch, antialias=0, extrapolation_value=0.0, exclude_outside=0, nearest_mode=floor, coordinate_transformation_mode=asymmetric, cubic_coeff_a=-0.75, mode=nearest}
     96 |  # node_cat
           %"cat"<FLOAT,[1,224,2*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),2*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Concat(%"upsample_nearest2d", %"max_pool2d_4") {axis=1}
     97 |  # node_convolution_6
           %"convolution_6"<FLOAT,[1,112,2*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),2*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Conv(%"cat", %"dec0.weight"{...}, %"dec0.bias"{...}) {group=1, pads=[1, 1, 1, 1], auto_pad=NOTSET, strides=[1, 1], dilations=[1, 1]}
     98 |  # node_relu_6
           %"relu_6"<FLOAT,[1,112,2*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),2*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Relu(%"convolution_6")
     99 |  # node_convolution_7
           %"convolution_7"<FLOAT,[1,112,2*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),2*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Conv(%"relu_6", %"con0.weight"{...}, %"con0.bias"{...}) {group=1, pads=[1, 1, 1, 1], auto_pad=NOTSET, strides=[1, 1], dilations=[1, 1]}
    100 |  # node_relu_7
           %"relu_7"<FLOAT,[1,112,2*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),2*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Relu(%"convolution_7")
    101 |  # node_Constant_18
           %"val_24"<?,?> ⬅️ ::Constant() {value_floats=[1.0, 1.0, 2.0, 2.0]}
    102 |  # node_upsample_nearest2d_1
           %"upsample_nearest2d_1"<FLOAT,[1,112,4*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),4*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Resize(%"relu_7", None, %"val_24") {keep_aspect_ratio_policy=stretch, antialias=0, extrapolation_value=0.0, exclude_outside=0, nearest_mode=floor, coordinate_transformation_mode=asymmetric, cubic_coeff_a=-0.75, mode=nearest}
    103 |  # node_cat_1
           %"cat_1"<FLOAT,[1,192,4*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),4*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Concat(%"upsample_nearest2d_1", %"max_pool2d_3") {axis=1}
    104 |  # node_convolution_8
           %"convolution_8"<FLOAT,[1,80,4*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),4*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Conv(%"cat_1", %"dec1.weight"{...}, %"dec1.bias"{...}) {group=1, pads=[1, 1, 1, 1], auto_pad=NOTSET, strides=[1, 1], dilations=[1, 1]}
    105 |  # node_relu_8
           %"relu_8"<FLOAT,[1,80,4*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),4*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Relu(%"convolution_8")
    106 |  # node_convolution_9
           %"convolution_9"<FLOAT,[1,80,4*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),4*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Conv(%"relu_8", %"con1.weight"{...}, %"con1.bias"{...}) {group=1, pads=[1, 1, 1, 1], auto_pad=NOTSET, strides=[1, 1], dilations=[1, 1]}
    107 |  # node_relu_9
           %"relu_9"<FLOAT,[1,80,4*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),4*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Relu(%"convolution_9")
    108 |  # node_Constant_19
           %"val_25"<?,?> ⬅️ ::Constant() {value_floats=[1.0, 1.0, 2.0, 2.0]}
    109 |  # node_upsample_nearest2d_2
           %"upsample_nearest2d_2"<FLOAT,[1,80,8*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),8*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Resize(%"relu_9", None, %"val_25") {keep_aspect_ratio_policy=stretch, antialias=0, extrapolation_value=0.0, exclude_outside=0, nearest_mode=floor, coordinate_transformation_mode=asymmetric, cubic_coeff_a=-0.75, mode=nearest}
    110 |  # node_cat_2
           %"cat_2"<FLOAT,[1,144,8*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),8*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Concat(%"upsample_nearest2d_2", %"max_pool2d_2") {axis=1}
    111 |  # node_convolution_10
           %"convolution_10"<FLOAT,[1,64,8*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),8*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Conv(%"cat_2", %"dec2.weight"{...}, %"dec2.bias"{...}) {group=1, pads=[1, 1, 1, 1], auto_pad=NOTSET, strides=[1, 1], dilations=[1, 1]}
    112 |  # node_relu_10
           %"relu_10"<FLOAT,[1,64,8*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),8*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Relu(%"convolution_10")
    113 |  # node_convolution_11
           %"convolution_11"<FLOAT,[1,64,8*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),8*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Conv(%"relu_10", %"con2.weight"{...}, %"con2.bias"{...}) {group=1, pads=[1, 1, 1, 1], auto_pad=NOTSET, strides=[1, 1], dilations=[1, 1]}
    114 |  # node_relu_11
           %"relu_11"<FLOAT,[1,64,8*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),8*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Relu(%"convolution_11")
    115 |  # node_Constant_20
           %"val_26"<?,?> ⬅️ ::Constant() {value_floats=[1.0, 1.0, 2.0, 2.0]}
    116 |  # node_upsample_nearest2d_3
           %"upsample_nearest2d_3"<FLOAT,[1,64,16*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),16*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Resize(%"relu_11", None, %"val_26") {keep_aspect_ratio_policy=stretch, antialias=0, extrapolation_value=0.0, exclude_outside=0, nearest_mode=floor, coordinate_transformation_mode=asymmetric, cubic_coeff_a=-0.75, mode=nearest}
    117 |  # node_cat_3
           %"cat_3"<FLOAT,[1,112,16*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),16*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Concat(%"upsample_nearest2d_3", %"max_pool2d_1") {axis=1}
    118 |  # node_convolution_12
           %"convolution_12"<FLOAT,[1,48,16*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),16*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Conv(%"cat_3", %"dec3.weight"{...}, %"dec3.bias"{...}) {group=1, pads=[1, 1, 1, 1], auto_pad=NOTSET, strides=[1, 1], dilations=[1, 1]}
    119 |  # node_relu_12
           %"relu_12"<FLOAT,[1,48,16*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),16*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Relu(%"convolution_12")
    120 |  # node_convolution_13
           %"convolution_13"<FLOAT,[1,48,16*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),16*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Conv(%"relu_12", %"con3.weight"{...}, %"con3.bias"{...}) {group=1, pads=[1, 1, 1, 1], auto_pad=NOTSET, strides=[1, 1], dilations=[1, 1]}
    121 |  # node_relu_13
           %"relu_13"<FLOAT,[1,48,16*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),16*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Relu(%"convolution_13")
    122 |  # node_Constant_21
           %"val_27"<?,?> ⬅️ ::Constant() {value_floats=[1.0, 1.0, 2.0, 2.0]}
    123 |  # node_upsample_nearest2d_4
           %"upsample_nearest2d_4"<FLOAT,[1,48,32*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),32*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Resize(%"relu_13", None, %"val_27") {keep_aspect_ratio_policy=stretch, antialias=0, extrapolation_value=0.0, exclude_outside=0, nearest_mode=floor, coordinate_transformation_mode=asymmetric, cubic_coeff_a=-0.75, mode=nearest}
    124 |  # node_cat_4
           %"cat_4"<FLOAT,[1,80,32*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),32*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Concat(%"upsample_nearest2d_4", %"max_pool2d") {axis=1}
    125 |  # node_convolution_14
           %"convolution_14"<FLOAT,[1,16,32*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),32*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Conv(%"cat_4", %"dec4.weight"{...}, %"dec4.bias"{...}) {group=1, pads=[1, 1, 1, 1], auto_pad=NOTSET, strides=[1, 1], dilations=[1, 1]}
    126 |  # node_relu_14
           %"relu_14"<FLOAT,[1,16,32*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),32*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Relu(%"convolution_14")
    127 |  # node_convolution_15
           %"convolution_15"<FLOAT,[1,16,32*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),32*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Conv(%"relu_14", %"con4.weight"{...}, %"con4.bias"{...}) {group=1, pads=[1, 1, 1, 1], auto_pad=NOTSET, strides=[1, 1], dilations=[1, 1]}
    128 |  # node_relu_15
           %"relu_15"<FLOAT,[1,16,32*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),32*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Relu(%"convolution_15")
    129 |  # node_Constant_22
           %"val_28"<?,?> ⬅️ ::Constant() {value_floats=[1.0, 1.0, 2.0, 2.0]}
    130 |  # node_upsample_nearest2d_5
           %"upsample_nearest2d_5"<FLOAT,[1,16,64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Resize(%"relu_15", None, %"val_28") {keep_aspect_ratio_policy=stretch, antialias=0, extrapolation_value=0.0, exclude_outside=0, nearest_mode=floor, coordinate_transformation_mode=asymmetric, cubic_coeff_a=-0.75, mode=nearest}
    131 |  # node_cat_5
           %"cat_5"<FLOAT,[1,19,64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Concat(%"upsample_nearest2d_5", %"pad") {axis=1}
    132 |  # node_convolution_16
           %"convolution_16"<FLOAT,[1,12,64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Conv(%"cat_5", %"dec5.weight"{...}, %"dec5.bias"{...}) {group=1, pads=[1, 1, 1, 1], auto_pad=NOTSET, strides=[1, 1], dilations=[1, 1]}
    133 |  # node_relu_16
           %"relu_16"<FLOAT,[1,12,64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Relu(%"convolution_16")
    134 |  # node_convolution_17
           %"convolution_17"<FLOAT,[1,12,64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Conv(%"relu_16", %"con5.weight"{...}, %"con5.bias"{...}) {group=1, pads=[1, 1, 1, 1], auto_pad=NOTSET, strides=[1, 1], dilations=[1, 1]}
    135 |  # node_relu_17
           %"relu_17"<FLOAT,[1,12,64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Relu(%"convolution_17")
    136 |  # node_Constant_23
           %"val_29"<?,?> ⬅️ ::Constant() {value_ints=[0]}
    137 |  # node_Constant_24
           %"val_30"<?,?> ⬅️ ::Constant() {value_ints=[9223372036854775807]}
    138 |  # node_Constant_25
           %"val_31"<?,?> ⬅️ ::Constant() {value=Tensor<INT64,[]>(array(0), name=None)}
    139 |  # node_Cast_26
           %"val_32"<?,?> ⬅️ ::Cast(%"val_31") {to=INT64}
    140 |  # node_Constant_27
           %"val_33"<?,?> ⬅️ ::Constant() {value_ints=[-1]}
    141 |  # node_Reshape_28
           %"val_34"<?,?> ⬅️ ::Reshape(%"val_32", %"val_33") {allowzero=0}
    142 |  # node_Constant_29
           %"val_35"<?,?> ⬅️ ::Constant() {value_ints=[1]}
    143 |  # node_slice_1
           %"slice_1"<FLOAT,[1,12,64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Slice(%"relu_17", %"val_29", %"val_30", %"val_34", %"val_35")
    144 |  # node_Constant_30
           %"val_36"<?,?> ⬅️ ::Constant() {value_ints=[0]}
    145 |  # node_Constant_31
           %"val_37"<?,?> ⬅️ ::Constant() {value_ints=[9223372036854775807]}
    146 |  # node_Cast_32
           %"val_38"<?,?> ⬅️ ::Cast(%"val_5") {to=INT64}
    147 |  # node_Constant_33
           %"val_39"<?,?> ⬅️ ::Constant() {value_ints=[-1]}
    148 |  # node_Reshape_34
           %"val_40"<?,?> ⬅️ ::Reshape(%"val_38", %"val_39") {allowzero=0}
    149 |  # node_Constant_35
           %"val_41"<?,?> ⬅️ ::Constant() {value_ints=[1]}
    150 |  # node_slice_2
           %"slice_2"<FLOAT,[1,12,64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64)),64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Slice(%"slice_1", %"val_36", %"val_37", %"val_40", %"val_41")
    151 |  # node_Constant_36
           %"val_42"<?,?> ⬅️ ::Constant() {value_ints=[0]}
    152 |  # node_Cast_37
           %"val_43"<?,?> ⬅️ ::Cast(%"sym_size_int_66") {to=INT64}
    153 |  # node_Constant_38
           %"val_44"<?,?> ⬅️ ::Constant() {value_ints=[-1]}
    154 |  # node_Reshape_39
           %"val_45"<?,?> ⬅️ ::Reshape(%"val_43", %"val_44") {allowzero=0}
    155 |  # node_Cast_40
           %"val_46"<?,?> ⬅️ ::Cast(%"val_9") {to=INT64}
    156 |  # node_Constant_41
           %"val_47"<?,?> ⬅️ ::Constant() {value_ints=[-1]}
    157 |  # node_Reshape_42
           %"val_48"<?,?> ⬅️ ::Reshape(%"val_46", %"val_47") {allowzero=0}
    158 |  # node_Constant_43
           %"val_49"<?,?> ⬅️ ::Constant() {value_ints=[1]}
    159 |  # node_slice_3
           %"slice_3"<FLOAT,[1,12,s95,64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))]> ⬅️ ::Slice(%"slice_2", %"val_42", %"val_45", %"val_48", %"val_49")
    160 |  # node_Constant_44
           %"val_50"<?,?> ⬅️ ::Constant() {value_ints=[0]}
    161 |  # node_Cast_45
           %"val_51"<?,?> ⬅️ ::Cast(%"sym_size_int_67") {to=INT64}
    162 |  # node_Constant_46
           %"val_52"<?,?> ⬅️ ::Constant() {value_ints=[-1]}
    163 |  # node_Reshape_47
           %"val_53"<?,?> ⬅️ ::Reshape(%"val_51", %"val_52") {allowzero=0}
    164 |  # node_Constant_48
           %"val_54"<?,?> ⬅️ ::Constant() {value=Tensor<INT64,[]>(array(3), name=None)}
    165 |  # node_Cast_49
           %"val_55"<?,?> ⬅️ ::Cast(%"val_54") {to=INT64}
    166 |  # node_Constant_50
           %"val_56"<?,?> ⬅️ ::Constant() {value_ints=[-1]}
    167 |  # node_Reshape_51
           %"val_57"<?,?> ⬅️ ::Reshape(%"val_55", %"val_56") {allowzero=0}
    168 |  # node_Constant_52
           %"val_58"<?,?> ⬅️ ::Constant() {value_ints=[1]}
    169 |  # node_slice_4
           %"output"<FLOAT,[1,12,s95,s28]> ⬅️ ::Slice(%"slice_3", %"val_50", %"val_53", %"val_57", %"val_58")
    return %"output"<FLOAT,[1,12,s95,s28]>
}

<
    opset_imports={'': 18},
>
def pkg.onnxscript.torch_lib.common::Rank(
    inputs=(
        %"input"<?,?>
    ),
    outputs=(
        %"return_val"<?,?>
    ),
) {
    0 |  # n0
         %"tmp"<?,?> ⬅️ ::Shape(%"input")
    1 |  # n1
         %"return_val"<?,?> ⬅️ ::Size(%"tmp")
    return %"return_val"<?,?>
}

<
    opset_imports={'': 18},
>
def pkg.onnxscript.torch_lib.common::IsScalar(
    inputs=(
        %"input"<?,?>
    ),
    outputs=(
        %"return_val"<?,?>
    ),
) {
    0 |  # n0
         %"tmp"<?,?> ⬅️ ::Shape(%"input")
    1 |  # n1
         %"tmp_0"<?,?> ⬅️ ::Size(%"tmp")
    2 |  # n2
         %"tmp_1"<?,?> ⬅️ ::Constant() {value_int=0}
    3 |  # n3
         %"return_val"<?,?> ⬅️ ::Equal(%"tmp_0", %"tmp_1")
    return %"return_val"<?,?>
}

Analysis

PyTorch ONNX Conversion Analysis

Model Information

The model has 1023692 parameters and 0 buffers (non-trainable parameters).
Number of parameters per dtype:

defaultdict(<class 'int'>, {torch.float32: 1023692})

Number of buffers per dtype:

defaultdict(<class 'int'>, {})

Inputs:

• i: TensorMetadata(shape=torch.Size([1, 3, s95, s28]), dtype=torch.float32, requires_grad=False, stride=(3*s28*s95, s28*s95, s28, 1), memory_format=torch.contiguous_format, is_quantized=False, qparams={})

Outputs:

• slice_4: TensorMetadata(shape=torch.Size([1, 12, s95, s28]), dtype=torch.float32, requires_grad=False, stride=(12*Max(1, 64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64)))*Max(1, 64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64))), Max(1, 64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64)))*Max(1, 64*(((s95 + (Mod(64 - (Mod(s95, 64)), 64)))//64))), Max(1, 64*(((s28 + (Mod(64 - (Mod(s28, 64)), 64)))//64))), 1), memory_format=torch.contiguous_format, is_quantized=False, qparams={})

The FX graph has 155 nodes in total. Number of FX nodes per op:

• placeholder: 37
• call_function: 117
• output: 1

Of the call_function nodes, the counts of operators used are:

• aten.convolution.default: 18
• aten.relu.default: 18
• <built-in function mul>: 14
• <built-in function floordiv>: 12
• <built-in function eq>: 12
• <built-in function ge>: 10
• aten.max_pool2d.default: 6
• aten.upsample_nearest2d.vec: 6
• aten.cat.default: 6
• <built-in function mod>: 4
• <built-in function add>: 4
• aten.slice.Tensor: 4
• aten.sym_size.int: 2
• aten.pad.default: 1

ONNX Conversion Information

All operators in the model have registered ONNX decompositions.

Decomposition comparison

Ops exist only in the ExportedProgram before decomposition: ['aten._assert_scalar.default', 'aten.conv2d.padding']

Ops exist only in the ExportedProgram after decomposition: ['aten.convolution.default']

[justinchuby]

Thanks. This is a pytorch issue. using DYNAMIC is the way to go.

[kistenklaus]

oki, thanks for maintaining +1