tico 0.1.0.dev250630__py3-none-any.whl → 0.1.0.dev250702__py3-none-any.whl

tico/__init__.py

@@ -21,7 +21,7 @@ from tico.config import CompileConfigV1, get_default_config
 from tico.utils.convert import convert, convert_from_exported_program, convert_from_pt2
 
 # THIS LINE IS AUTOMATICALLY GENERATED BY setup.py
-__version__ = "0.1.0.dev250630"
+__version__ = "0.1.0.dev250702"
 
 MINIMUM_SUPPORTED_VERSION = "2.5.0"
 SECURE_TORCH_VERSION = "2.6.0"

tico/interpreter/infer.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Any
+from typing import Any, Sequence
 
 import numpy as np
 import torch
@@ -20,40 +20,46 @@ from circle_schema import circle
 
 from tico.interpreter.interpreter import Interpreter
 from tico.serialize.circle_mapping import np_dtype_from_circle_dtype, to_circle_dtype
+from tico.utils.installed_packages import is_dynamic_cache_available
 
 
-def preprocess_inputs(inputs: Any):
-    """
-    Preprocess user inputs for circle inference.
+def flatten_and_convert(inputs: Sequence) -> tuple:
+    result = []  # type: ignore[var-annotated]
+    for item in inputs:
+        if item is None:
+            continue
 
-    1. None inputs are ignored.
-    2. A list/tuple input is flatten when a torch module is exported.
-      e.g. inputs = (torch.Tensor, [2,3,4]) -> inputs = (torch.Tensor, 2, 3, 4)
-    """
-    l = []
-    for value in inputs:
-        if value == None:
+        # 1. recursion on list and tuple
+        if isinstance(item, (list, tuple)):
+            result.extend(flatten_and_convert(item))
             continue
-        if isinstance(value, (tuple, list)):
-            for val in value:
-                l.append(val)
-        else:
-            l.append(value)
-    # Check if it is a list of a list.
-    if any(isinstance(item, (tuple, list)) for item in l):
-        l = preprocess_inputs(l)
-    return tuple(l)
+
+        # 2. handle DynamicCache
+        if is_dynamic_cache_available():
+            from transformers.cache_utils import DynamicCache
+
+            if isinstance(item, DynamicCache):
+                # NOTE The tensor order is: key_in → key_out → value_in → value_out
+                #
+                # Refer to https://github.com/huggingface/transformers/blob/3457e8e73e4f5532cc69059682b1ba4484d7e7e8/src/transformers/cache_utils.py#L557
+                # ```py
+                # self.key_cache[layer_idx] = torch.cat([self.key_cache[layer_idx], key_states], dim=-2)
+                # self.value_cache[layer_idx] = torch.cat([self.value_cache[layer_idx], value_states], dim=-2)
+                # ```
+                result.extend(item.key_cache)
+                result.extend(item.value_cache)
+                continue
+
+        # 3. Convert to tensors
+        result.append(item if isinstance(item, torch.Tensor) else torch.tensor(item))
+
+    return tuple(result)
 
 
 def infer(circle_binary: bytes, *args: Any, **kwargs: Any) -> Any:
     # When converting a model, it is assumed that the order of keyword arguments is maintained.
-    user_inputs = args + tuple(kwargs.values())
-    user_inputs = preprocess_inputs(user_inputs)
-    # Cast them to torch.Tensor to make it simple.
-    user_inputs = tuple(
-        torch.tensor(user_input) if type(user_input) != torch.Tensor else user_input
-        for user_input in user_inputs
-    )
+    raw_inputs = args + tuple(kwargs.values())
+    user_inputs = flatten_and_convert(raw_inputs)
 
     # Get input spec from circle binary.
     model = circle.Model.Model.GetRootAsModel(circle_binary, 0)

tico/passes/legalize_predefined_layout_operators.py

@@ -30,6 +30,7 @@ from tico.utils.utils import is_target_node
 from tico.utils.validate_args_kwargs import (
     AvgPool2dArgs,
     Conv2DArgs,
+    ConvTranspose2DArgs,
     DequantizePerChannelArgs,
     DequantizePerTensorArgs,
     InstanceNormArgs,
@@ -37,7 +38,9 @@ from tico.utils.validate_args_kwargs import (
 )
 
 
-def get_permute_weight_input(conv_args: Conv2DArgs) -> torch.fx.Node:
+def get_permute_weight_input(
+    conv_args: Conv2DArgs | ConvTranspose2DArgs,
+) -> torch.fx.Node:
     """
     Retrieves the weight input for the permute operation.
 
@@ -194,6 +197,85 @@ class LegalizePreDefinedLayoutOperators(PassBase):
         modified = True
         return modified
 
+    def legalize_conv_transpose2d(self, exported_program, node) -> bool:
+        logger = logging.getLogger(__name__)
+        modified = False
+
+        graph_module = exported_program.graph_module
+        graph = graph_module.graph
+
+        args = ConvTranspose2DArgs(*node.args, **node.kwargs)  # type: ignore[arg-type]
+        input = args.input
+        padding = args.padding
+        groups = args.groups
+        dilation = args.dilation
+
+        input_shape = extract_shape(input)
+        if not (len(input_shape) == 4):
+            raise NotYetSupportedError(
+                f"Only support 4D input tensor: node's input shape: {input_shape}"
+            )
+
+        if groups != 1:
+            raise NotYetSupportedError(
+                f"Only support groups=1: node's groups: {groups}"
+            )
+
+        if dilation != [1, 1]:
+            raise NotYetSupportedError(
+                f"Only support dilation=[1, 1]: node's groups: {dilation}"
+            )
+
+        NCHW_to_NHWC = [0, 2, 3, 1]
+        # input permute
+        with graph.inserting_after(input):
+            input_permute = create_node(
+                graph,
+                torch.ops.aten.permute.default,
+                args=(input, NCHW_to_NHWC),
+                origin=input,
+            )
+            node.update_arg(node.args.index(input), input_permute)
+
+        # weight permute
+        weight = get_permute_weight_input(args)
+        with graph.inserting_after(weight):
+            perm = [1, 2, 3, 0]  # IOHW_to_OHWI
+            weight_permute = create_node(
+                graph,
+                torch.ops.aten.permute.default,
+                args=(weight, perm),
+                origin=weight,
+            )
+            if args.weight.target in [
+                torch.ops.quantized_decomposed.dequantize_per_channel.default,
+                torch.ops.quantized_decomposed.dequantize_per_tensor.default,
+            ]:
+                dq = args.weight
+                dq.update_arg(dq.args.index(weight), weight_permute)
+                # Need to update dq.meta["val"] in FillMetaVal pass.
+                del dq.meta["val"]
+            else:
+                node.update_arg(node.args.index(weight), weight_permute)
+
+        with graph.inserting_before(node):
+            legalized_op = torch.ops.circle_custom.transpose_conv
+            circle_op = create_node(
+                graph, legalized_op, args=node.args, kwargs=node.kwargs, origin=node
+            )
+            # output permute
+            NHWC_to_NCHW = [0, 3, 1, 2]
+            conv_out_permute = create_node(
+                graph,
+                torch.ops.aten.permute.default,
+                args=(circle_op, NHWC_to_NCHW),
+            )
+        node.replace_all_uses_with(conv_out_permute, propagate_meta=True)
+
+        logger.debug(f"{node.name} is replaced with {circle_op.name}")
+        modified = True
+        return modified
+
     def legalize_instance_norm(self, exported_program, node) -> bool:
         logger = logging.getLogger(__name__)
         modified = False
@@ -365,6 +447,7 @@ class LegalizePreDefinedLayoutOperators(PassBase):
         target_to_legalize_func = {
             torch.ops.aten.conv2d.default: self.legalize_conv2d,
             torch.ops.aten.conv2d.padding: self.legalize_conv2d,
+            torch.ops.aten.conv_transpose2d.input: self.legalize_conv_transpose2d,
             torch.ops.aten.max_pool2d_with_indices.default: self.legalize_max_pool2d_with_indices,
             torch.ops.aten.avg_pool2d.default: self.legalize_avg_pool2d,
             torch.ops.aten.instance_norm.default: self.legalize_instance_norm,

tico/serialize/operators/op_conv2d.py

@@ -26,7 +26,7 @@ from tico.serialize.operators.node_visitor import NodeVisitor, register_node_vis
 from tico.serialize.operators.utils import create_builtin_operator, get_op_index
 from tico.serialize.quant_param import QPARAM_KEY, QuantParam
 from tico.utils.define import define_pad_node
-from tico.utils.padding import is_same_padding, is_valid_padding, SAME, VALID
+from tico.utils.padding import identify_padding
 from tico.utils.validate_args_kwargs import Conv2DArgs
 
 
@@ -111,53 +111,39 @@ class Conv2dVisitor(NodeVisitor):
 
         assert groups == 1, "Only support group 1 conv2d"
 
-        input_dtype: int = extract_circle_dtype(input_)
         input_shape = list(extract_shape(input_))
+        output_shape = list(extract_shape(node))
+        weight_shape = list(extract_shape(weight))
         assert len(input_shape) == 4, len(input_shape)
-        output_shape = extract_shape(node)
         assert len(output_shape) == 4, len(output_shape)
+        assert len(weight_shape) == 4, len(weight_shape)
 
-        conv_input: torch.fx.node.Node | circle.Tensor.TensorT = input_
-        weight_shape = list(extract_shape(weight))
-
-        if is_valid_padding(padding):
-            conv2d_padding_type = VALID
-        elif is_same_padding(padding, input_shape, output_shape) and stride == [1, 1]:
-            conv2d_padding_type = SAME
-        else:
-            assert isinstance(padding, list) and len(padding) == 2
+        pad_decision = identify_padding(padding, input_shape, output_shape, stride)
 
-            conv2d_padding_type = VALID
-
-            # Padding is not valid or same, so we use valid padding and add padding operator before conv2d operator.
-            # when data_foramt is "NHWC", padding should be [[0, 0], [pad_top, pad_bottom], [pad_left, pad_right], [0, 0]]
+        conv_input: torch.fx.Node | circle.Tensor.TensorT = input_
+        if pad_decision.explicit_pad_hw is not None:
+            pad_h, pad_w = pad_decision.explicit_pad_hw
             paddings = torch.tensor(
                 [
                     [0, 0],
-                    [padding[0], padding[0]],
-                    [padding[1], padding[1]],
+                    [pad_h, pad_h],
+                    [pad_w, pad_w],
                     [0, 0],
                 ],
                 dtype=torch.int32,
             )
             pad_output_shape = [
                 input_shape[0],
-                input_shape[1],
-                input_shape[2],
+                input_shape[1] + pad_h * 2,
+                input_shape[2] + pad_w * 2,
                 input_shape[3],
             ]
-            # Add (pad_top+pad_bottom) to pad_output_shape_h
-            pad_output_shape[1] += padding[0] * 2
-            # Add (pad_left+pad_Right) to pad_output_shape_w
-            pad_output_shape[2] += padding[1] * 2
             # create padded output tensor
-            input_qparam: Optional[QuantParam] = (
-                input_.meta[QPARAM_KEY] if QPARAM_KEY in input_.meta else None
-            )
+            input_qparam: Optional[QuantParam] = input_.meta.get(QPARAM_KEY)
             pad_output = self.graph.add_tensor_from_scratch(
                 prefix=f"{node.name}_input_pad_output",
                 shape=pad_output_shape,
-                dtype=input_dtype,
+                dtype=extract_circle_dtype(input_),
                 qparam=input_qparam,
                 source_node=node,
             )
@@ -170,13 +156,11 @@ class Conv2dVisitor(NodeVisitor):
 
         if bias is None:
             # luci-interpreter can't run no bias conv. Let's add zero vector for bias.
-            assert len(weight_shape) == 4
-            out_channel = weight_shape[0]
-            bias = [0.0] * out_channel  # type: ignore[assignment]
+            bias = [0.0] * weight_shape[0]  # type: ignore[assignment]
 
         # Conv2D
         conv2d_operator = self.define_conv2d_node(
-            conv2d_padding_type,  # 'SAME'(0) or 'VALID'(1)
+            pad_decision.conv_padding_type,  # 'SAME'(0) or 'VALID'(1)
             stride,
             dilation,
             [conv_input, weight, bias],

tico/serialize/operators/op_depthwise_conv2d.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Dict, List, TYPE_CHECKING
+from typing import Dict, List, Optional, TYPE_CHECKING
 
 if TYPE_CHECKING:
     import torch._ops
@@ -24,8 +24,9 @@ from tico.serialize.circle_mapping import extract_circle_dtype, extract_shape
 from tico.serialize.operators.hashable_opcode import OpCode
 from tico.serialize.operators.node_visitor import NodeVisitor, register_node_visitor
 from tico.serialize.operators.utils import create_builtin_operator, get_op_index
+from tico.serialize.quant_param import QPARAM_KEY, QuantParam
 from tico.utils.define import define_pad_node
-from tico.utils.padding import is_same_padding, is_valid_padding, SAME, VALID
+from tico.utils.padding import identify_padding
 from tico.utils.validate_args_kwargs import Conv2DArgs
 
 
@@ -114,63 +115,49 @@ class DepthwiseConv2dVisitor(NodeVisitor):
         dilation = args.dilation
         groups = args.groups
 
-        input_dtype: int = extract_circle_dtype(input_)
         input_shape = list(extract_shape(input_))  # OHWI
-        assert len(input_shape) == 4, len(input_shape)
-
         output_shape = list(extract_shape(node))  # OHWI
-        assert len(output_shape) == 4, len(output_shape)
-
         weight_shape = list(extract_shape(weight))  # 1HWO
-        assert (
-            weight_shape[3] % groups == 0
-        ), "Depthwise convolution requires output channel to be divisible by groups"
-
+        assert len(input_shape) == 4, len(input_shape)
+        assert len(output_shape) == 4, len(output_shape)
+        assert len(weight_shape) == 4
         assert weight_shape[0] == 1
         assert weight_shape[3] == output_shape[3]
         assert input_shape[3] == groups
+        assert (
+            weight_shape[3] % groups == 0
+        ), "Depthwise convolution requires output channel to be divisible by groups"
 
         depthMultiplier = weight_shape[3] // input_shape[3]
         assert weight_shape[3] % input_shape[3] == 0, "depthMultiplier must be integer"
 
-        conv_input: torch.fx.node.Node | circle.Tensor.TensorT = input_
-
-        if is_valid_padding(padding):
-            dconv2d_padding_type = VALID
-        elif is_same_padding(padding, input_shape, output_shape):
-            dconv2d_padding_type = SAME
-        else:
-            assert isinstance(padding, list) and len(padding) == 2
+        pad_decision = identify_padding(padding, input_shape, output_shape, stride)
 
-            dconv2d_padding_type = VALID
-
-            # Padding is not valid or same, so we use valid padding and add padding operator before conv2d operator.
-            # when data_format is "NHWC", padding should be [[0, 0], [pad_top, pad_bottom], [pad_left, pad_right], [0, 0]]
+        conv_input: torch.fx.Node | circle.Tensor.TensorT = input_
+        if pad_decision.explicit_pad_hw is not None:
+            pad_h, pad_w = pad_decision.explicit_pad_hw
             paddings = torch.tensor(
                 [
                     [0, 0],
-                    [padding[0], padding[0]],
-                    [padding[1], padding[1]],
+                    [pad_h, pad_h],
+                    [pad_w, pad_w],
                     [0, 0],
                 ],
                 dtype=torch.int32,
             )
             pad_output_shape = [
                 input_shape[0],
-                input_shape[1],
-                input_shape[2],
+                input_shape[1] + pad_h * 2,
+                input_shape[2] + pad_w * 2,
                 input_shape[3],
             ]
-            # Add (pad_top+pad_bottom) to pad_output_shape_h
-            pad_output_shape[1] += padding[0] * 2
-            # Add (pad_left+pad_Right) to pad_output_shape_w
-            pad_output_shape[2] += padding[1] * 2
             # create padded output tensor
-
+            input_qparam: Optional[QuantParam] = input_.meta.get(QPARAM_KEY)
             pad_output = self.graph.add_tensor_from_scratch(
                 prefix=f"{node.name}_input_pad_output",
                 shape=pad_output_shape,
-                dtype=input_dtype,
+                dtype=extract_circle_dtype(input_),
+                qparam=input_qparam,
                 source_node=node,
             )
             # CirclePad
@@ -182,13 +169,11 @@ class DepthwiseConv2dVisitor(NodeVisitor):
 
         if bias is None:
             # luci-interpreter can't run no bias conv. Let's add zero vector for bias.
-            assert len(weight_shape) == 4
-            out_channel = weight_shape[3]
-            bias = [0.0] * out_channel  # type: ignore[assignment]
+            bias = [0.0] * weight_shape[3]  # type: ignore[assignment]
 
         # DConv2D
         dconv2d_operator = self.define_dconv_node(
-            dconv2d_padding_type,
+            pad_decision.conv_padding_type,
             stride,
             dilation,
             depthMultiplier,

tico/serialize/operators/op_transpose_conv.py

@@ -0,0 +1,149 @@
+# Copyright (c) 2025 Samsung Electronics Co., Ltd. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Dict, List, Optional, TYPE_CHECKING
+
+if TYPE_CHECKING:
+    import torch._ops
+    import torch.fx
+import torch
+from circle_schema import circle
+
+from tico.serialize.circle_mapping import (
+    circle_legalize_dtype_to,
+    extract_circle_dtype,
+    extract_shape,
+)
+from tico.serialize.operators.hashable_opcode import OpCode
+from tico.serialize.operators.node_visitor import NodeVisitor, register_node_visitor
+from tico.serialize.operators.utils import create_builtin_operator, get_op_index
+from tico.serialize.quant_param import QPARAM_KEY, QuantParam
+from tico.utils.define import define_pad_node
+from tico.utils.padding import identify_padding
+from tico.utils.validate_args_kwargs import ConvTranspose2DArgs
+
+
+@register_node_visitor
+class TransposeConvVisitor(NodeVisitor):
+    target: List[torch._ops.OpOverload] = [
+        torch.ops.circle_custom.transpose_conv,
+    ]
+
+    def define_transpose_conv_node(
+        self, padding: int, stride: List, inputs: List, outputs: List
+    ) -> circle.Operator.OperatorT:
+        def set_transpose_conv_option(operator, stride):
+            operator.builtinOptionsType = (
+                circle.BuiltinOptions.BuiltinOptions.TransposeConvOptions
+            )
+            option = circle.TransposeConvOptions.TransposeConvOptionsT()
+            option.padding = padding
+            option.strideH = stride[0]
+            option.strideW = stride[1]
+            option.fusedActivationFunction = (
+                circle.ActivationFunctionType.ActivationFunctionType.NONE
+            )
+            operator.builtinOptions = option
+
+        transpose_conv_op_index = get_op_index(
+            circle.BuiltinOperator.BuiltinOperator.TRANSPOSE_CONV, self._op_codes
+        )
+        operator = create_builtin_operator(
+            self.graph, transpose_conv_op_index, inputs, outputs
+        )
+        set_transpose_conv_option(operator, stride)
+        return operator
+
+    def __init__(self, op_codes: Dict[OpCode, int], graph):
+        super().__init__(op_codes, graph)
+
+    def define_node(self, node: torch.fx.Node) -> circle.Operator.OperatorT:
+        args = ConvTranspose2DArgs(*node.args, **node.kwargs)  # type: ignore[arg-type]
+
+        input_ = args.input
+        weight = args.weight
+        bias = args.bias
+        stride = args.stride
+        padding = args.padding
+        output_padding = args.output_padding
+        groups = args.groups
+        dilation = args.dilation
+
+        assert groups == 1, "Only support group 1"
+
+        input_shape = list(extract_shape(input_))
+        output_shape = list(extract_shape(node))
+        weight_shape = list(extract_shape(weight))
+        assert len(input_shape) == 4, len(input_shape)
+        assert len(output_shape) == 4, len(output_shape)
+        assert len(weight_shape) == 4, len(weight_shape)
+
+        pad_decision = identify_padding(padding, input_shape, output_shape, stride)
+
+        conv_input: torch.fx.Node | circle.Tensor.TensorT = input_
+        if pad_decision.explicit_pad_hw is not None:
+            pad_h, pad_w = pad_decision.explicit_pad_hw
+            paddings = torch.tensor(
+                [
+                    [0, 0],
+                    [pad_h, pad_h],
+                    [pad_w, pad_w],
+                    [0, 0],
+                ],
+                dtype=torch.int32,
+            )
+            pad_output_shape = [
+                input_shape[0],
+                input_shape[1] + pad_h * 2,
+                input_shape[2] + pad_w * 2,
+                input_shape[3],
+            ]
+            # create padded output tensor
+            input_qparam: Optional[QuantParam] = input_.meta.get(QPARAM_KEY)
+            pad_output = self.graph.add_tensor_from_scratch(
+                prefix=f"{node.name}_input_pad_output",
+                shape=pad_output_shape,
+                dtype=extract_circle_dtype(input_),
+                qparam=input_qparam,
+                source_node=node,
+            )
+            # CirclePad
+            pad_operator = define_pad_node(
+                self.graph, self._op_codes, [input_, paddings], [pad_output]
+            )
+            self.graph.add_operator(pad_operator)
+            conv_input = pad_output
+
+        if bias is None:
+            # luci-interpreter can't run no bias conv. Let's add zero vector for bias.
+            bias = [0.0] * weight_shape[0]  # type: ignore[assignment]
+
+        # First arguemnt is output shape of tconv.
+        assert output_shape[0] == input_shape[0]
+        assert output_shape[3] == weight_shape[0]
+        tconv_output = circle_legalize_dtype_to(output_shape, dtype=torch.int32)
+
+        tconv_output_tensor = self.graph.add_const_tensor(
+            tconv_output, source_node=node
+        )
+
+        # TConv2D
+        tconv2d_operator = self.define_transpose_conv_node(
+            pad_decision.conv_padding_type,  # 'SAME'(0) or 'VALID'(1)
+            stride,
+            [tconv_output_tensor, weight, conv_input, bias],
+            [node],
+        )
+
+        return tconv2d_operator

tico/utils/convert.py

@@ -100,6 +100,7 @@ def traced_run_decompositions(exported_program: ExportedProgram):
             torch.ops.aten.conv2d.padding,
             torch.ops.aten.conv1d.default,
             torch.ops.aten.conv1d.padding,
+            torch.ops.aten.conv_transpose2d.input,
             torch.ops.aten.instance_norm.default,
             torch.ops.aten._safe_softmax.default,
             torch.ops.aten.relu6.default,  # Do not decompose to hardtanh
@@ -116,6 +117,7 @@ def traced_run_decompositions(exported_program: ExportedProgram):
             torch.ops.aten.conv2d.padding,
             torch.ops.aten.conv1d.default,
             torch.ops.aten.conv1d.padding,
+            torch.ops.aten.conv_transpose2d.input,
             torch.ops.aten.instance_norm.default,
             torch.ops.aten._safe_softmax.default,
             torch.ops.aten.relu6.default,  # Do not decompose to hardtanh
@@ -287,6 +289,12 @@ def convert(
     strict: bool = True,
     config: CompileConfigBase = get_default_config(),
 ) -> CircleModel:
+    if hasattr(mod, "training") and mod.training:
+        logger = logging.getLogger(__name__)
+        logger.fatal(
+            "Your model is in TRAINING MODE. PLEASE CHECK IF YOU FORGOT `model.eval()`."
+        )
+
     with torch.no_grad():
         exported_program = export(mod, args, kwargs, strict=strict)
 

tico/utils/installed_packages.py

@@ -0,0 +1,35 @@
+# Copyright (c) 2025 Samsung Electronics Co., Ltd. All Rights Reserved
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+##############################
+#### Transformers Package ####
+##############################
+
+
+def is_transformers_installed():
+    try:
+        import transformers
+
+        return True
+    except ImportError:
+        return False
+
+
+def is_dynamic_cache_available():
+    try:
+        from transformers.cache_utils import DynamicCache
+
+        return True
+    except ImportError:
+        return False

tico/utils/padding.py

@@ -12,36 +12,71 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from enum import IntEnum
+from typing import NamedTuple, Optional, Sequence, Tuple, Union
+
 import torch
 
 from tico.utils.errors import InvalidArgumentError
 
-SAME = 0
-VALID = 1
 
+PaddingValue = Union[str, Sequence[int]]  # "same" | "valid" | [pad_h, pad_w]
 
-def is_valid_padding(padding: str | list):
-    if isinstance(padding, str):
-        return padding == "valid"
 
-    if isinstance(padding, list):
-        assert len(padding) == 2, "Padding should be a list of length 2."
-        return padding == [0, 0]
+class ConvPadding(IntEnum):
+    SAME = 0  # auto-pad, HW out == HW in
+    VALID = 1  # no implicit padding
+
 
-    raise InvalidArgumentError("Invalid padding.")
+class ConvPaddingInfo(NamedTuple):
+    """
+    Result of padding analysis.
+    """
 
+    conv_padding_type: ConvPadding
+    explicit_pad_hw: Optional[Tuple[int, int]]  # None -> no extra Pad() op needed
 
-def is_same_padding(
-    padding: str | list, input_shape: list | torch.Size, output_shape: list | torch.Size
-):
+
+def identify_padding(
+    padding: PaddingValue,
+    input_shape: Sequence[int],
+    output_shape: Sequence[int],
+    stride: Sequence[int],
+) -> ConvPaddingInfo:
+    """
+    Normalizes all PyTorch `padding` variants to a single decision.
+
+    Rules
+    -----
+    1. "valid" or [0, 0]                              → VALID, no Pad().
+    2. "same" or the shapes already match (stride==1) → SAME, no Pad().
+    3. Any other 2-element list                       → VALID + explicit Pad().
+    """
+    # ─── 1. String form ────────────────────────────────────────────────────
     if isinstance(padding, str):
-        return padding == "same"
+        pad = padding.lower()
+        if pad == "valid":
+            return ConvPaddingInfo(ConvPadding.VALID, None)
+        if pad == "same":
+            return ConvPaddingInfo(ConvPadding.SAME, None)
+        raise InvalidArgumentError(f"Unknown padding string: {padding}")
+
+    # ─── 2. List / tuple form ─────────────────────────────────────────────
+    if not (isinstance(padding, (list, tuple)) and len(padding) == 2):
+        raise InvalidArgumentError(
+            "Padding must be 'valid', 'same', or a [pad_h, pad_w] list"
+        )
 
-    if isinstance(padding, list):
-        assert len(padding) == 2, "Padding should be a list of length 2."
+    pad_h, pad_w = padding
+    # [0, 0]  → VALID
+    if pad_h == 0 and pad_w == 0:
+        return ConvPaddingInfo(ConvPadding.VALID, None)
 
-        input_HW = tuple(input_shape[1:3])  # N H W C
-        output_HW = tuple(output_shape[1:3])  # N H W C
-        return input_HW == output_HW
+    # SAME heuristic: output H/W already match input when stride is 1
+    hw_in = tuple(input_shape[1:3])
+    hw_out = tuple(output_shape[1:3])
+    if hw_in == hw_out and stride == [1, 1]:
+        return ConvPaddingInfo(ConvPadding.SAME, None)
 
-    raise InvalidArgumentError("Invalid padding.")
+    # Anything else = explicit symmetric padding
+    return ConvPaddingInfo(ConvPadding.VALID, (pad_h, pad_w))

tico/utils/register_custom_op.py

@@ -371,6 +371,113 @@ def CircleDepthwiseConv2dPadding():
         return NHWC_output
 
 
+def CircleTransposeConv():
+    """
+    Note that this op follows the input spec of `aten.conv_transpose2d.input` whose number
+     of arguments meets (2 <= node.args <= 8) condition.
+    [RESTRICTION]
+      Therefore, I tried to define a spec of it as transpose_conv(input, weight, *args).
+      But, custom operators in torch do not support positional-only args. So, I set it
+       them as None by default.
+    """
+
+    @custom_op("circle_custom::transpose_conv", mutates_args=())
+    def transpose_conv(
+        input_: torch.Tensor,
+        weight: torch.Tensor,
+        bias: Optional[torch.Tensor] = None,
+        stride: Optional[List[int]] = None,
+        padding: Optional[List[int]] = None,
+        output_padding: Optional[List[int]] = None,
+        groups: Optional[int] = None,
+        dilation: Optional[List[int]] = None,
+    ) -> torch.Tensor:
+        """
+        Set default values.
+        Custom operators have limited types when it comes to default values.
+        So, let's set them by None in input specs, and then, set it by default values.
+        https://github.com/pytorch/pytorch/blob/6b05aafc/torch/_library/infer_schema.py#L131-L144
+        """
+        stride = [1, 1] if stride is None else stride
+        padding = [0, 0] if padding is None else padding
+        output_padding = [0, 0] if output_padding is None else output_padding
+        groups = 1 if groups is None else groups
+        dilation = [1, 1] if dilation is None else dilation
+        if groups != 1:
+            raise RuntimeError(
+                f"CircleTransposeConv only supports 1 'groups'. the node's groups: {groups}"
+            )
+
+        NHWC_to_NCHW = [0, 3, 1, 2]
+        OHWI_to_IOHW = [3, 0, 1, 2]
+        NCHW_input = torch.ops.aten.permute.default(input_, NHWC_to_NCHW)
+        OIHW_weight = torch.ops.aten.permute.default(weight, OHWI_to_IOHW)
+
+        args = [
+            NCHW_input,
+            OIHW_weight,
+            bias,
+            stride,
+            padding,
+            output_padding,
+            groups,
+            dilation,
+        ]
+        NCHW_output = torch.ops.aten.conv_transpose2d.input(*args)
+        NCHW_to_NHWC = [0, 2, 3, 1]
+        NHWC_output = torch.ops.aten.permute.default(NCHW_output, NCHW_to_NHWC)
+
+        return NHWC_output
+
+    @register_fake("circle_custom::transpose_conv")
+    def _(
+        input_: torch.Tensor,
+        weight: torch.Tensor,
+        bias: Optional[torch.Tensor] = None,
+        stride: Optional[List[int]] = None,
+        padding: Optional[List[int]] = None,
+        output_padding: Optional[List[int]] = None,
+        groups: Optional[int] = None,
+        dilation: Optional[List[int]] = None,
+    ):
+        """
+        Set default values.
+        Custom operators have limited types when it comes to default values.
+        So, let's set them by None in input specs, and then, set it by default values.
+        https://github.com/pytorch/pytorch/blob/6b05aafc/torch/_library/infer_schema.py#L131-L144
+        """
+        stride = [1, 1] if stride is None else stride
+        padding = [0, 0] if padding is None else padding
+        output_padding = [0, 0] if output_padding is None else output_padding
+        groups = 1 if groups is None else groups
+        dilation = [1, 1] if dilation is None else dilation
+        if groups != 1:
+            raise RuntimeError(
+                f"CircleConv2d only supports 1 'groups'. the node's groups: {groups}"
+            )
+
+        NHWC_to_NCHW = [0, 3, 1, 2]
+        OHWI_to_IOHW = [3, 0, 1, 2]
+        NCHW_input = torch.ops.aten.permute.default(input_, NHWC_to_NCHW)
+        OIHW_weight = torch.ops.aten.permute.default(weight, OHWI_to_IOHW)
+
+        args = [
+            NCHW_input,
+            OIHW_weight,
+            bias,
+            stride,
+            padding,
+            output_padding,
+            groups,
+            dilation,
+        ]
+        NCHW_output = torch.ops.aten.conv_transpose2d.input(*args)
+        NCHW_to_NHWC = [0, 2, 3, 1]
+        NHWC_output = torch.ops.aten.permute.default(NCHW_output, NCHW_to_NHWC)
+
+        return NHWC_output
+
+
 def CircleMaxPool2D():
     """
     Note that this op follows the input spec of `aten.max_pool2d_with_indices.default` whose number
@@ -603,6 +710,7 @@ def RegisterOps():
     CircleDepthwiseConv2dPadding()
     CircleConv2d()
     CircleConv2dPadding()
+    CircleTransposeConv()
     CircleMaxPool2D()
     CircleAvgPool2D()
     CircleInstanceNorm()

tico/utils/validate_args_kwargs.py

@@ -208,6 +208,45 @@ class ConstantPadNdArgs:
     value: int | float
 
 
+@enforce_type
+@dataclass
+class ConvArgs:
+    """
+    convolution(Tensor input, Tensor weight, Tensor? bias, SymInt[] stride, SymInt[] padding, SymInt[] dilation, bool transposed, SymInt[] output_padding, SymInt groups) -> Tensor
+    """
+
+    input: torch.fx.Node
+    weight: torch.fx.Node
+    bias: Union[torch.fx.Node, None]
+    stride: List[int]
+    padding: List[int]
+    dilation: List[int]
+    transposed: bool
+    output_padding: List[int]
+    groups: int
+
+
+@enforce_type
+@dataclass
+class ConvTranspose2DArgs:
+    """
+    conv_transpose2d.input(Tensor input, Tensor weight, Tensor? bias=None, SymInt[2] stride=1, SymInt[2] padding=0, SymInt[2] output_padding=0, SymInt groups=1, SymInt[2] dilation=1) -> Tensor
+    """
+
+    input: torch.fx.Node
+    weight: torch.fx.Node
+    bias: Union[torch.fx.Node, None] = None
+    stride: List[int] = field(default_factory=lambda: [1, 1])
+    padding: List[int] = field(default_factory=lambda: [0, 0])
+    output_padding: List[int] = field(default_factory=lambda: [0, 0])
+    groups: int = 1
+    dilation: List[int] = field(default_factory=lambda: [1, 1])
+
+    def __post_init__(self):
+        assert len(self.stride) == 2, len(self.stride)
+        assert len(self.dilation) == 2, len(self.dilation)
+
+
 @enforce_type
 @dataclass
 class Conv2DArgs:

{tico-0.1.0.dev250630.dist-info → tico-0.1.0.dev250702.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: tico
-Version: 0.1.0.dev250630
+Version: 0.1.0.dev250702
 Summary: Convert exported Torch module to circle
 Home-page: UNKNOWN
 License: UNKNOWN

{tico-0.1.0.dev250630.dist-info → tico-0.1.0.dev250702.dist-info}/RECORD

@@ -1,4 +1,4 @@
-tico/__init__.py,sha256=gNG4gqEHE73hnLBf47wz6Gj7RXqTV5DoQ8Bpss9Az84,1743
+tico/__init__.py,sha256=OwKeuduBe-XLK_9XzKLwmCWJd6kRwEZ2xiY8RnRGnCA,1743
 tico/pt2_to_circle.py,sha256=gu3MD4Iqc0zMZcCZ2IT8oGbyj21CTSbT3Rgd9s2B_9A,2767
 tico/config/__init__.py,sha256=xZzCXjZ84qE-CsBi-dfaL05bqpQ3stKKfTXhnrJRyVs,142
 tico/config/base.py,sha256=anwOiJFkUxUi7Cef573JgQcjk6S-FSi6O_TLjYASW-g,1244
@@ -57,7 +57,7 @@ tico/experimental/quantization/passes/propagate_qparam_forward.py,sha256=RhUHGCR
 tico/experimental/quantization/passes/quantize_bias.py,sha256=ZQ3rETYStpW28JUbODRixbq5sDEOiIOB_qWA-Jzuu-Y,4337
 tico/experimental/quantization/passes/remove_weight_dequant_op.py,sha256=Klc_9-94tl0_AuAToKOjsWED_YPk5RB67eum0ddPX7o,6588
 tico/interpreter/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
-tico/interpreter/infer.py,sha256=vJ3b69ce9HrxNT0gFwbEhHpAyvVyuiunTgAeiqn5t64,4350
+tico/interpreter/infer.py,sha256=1ZFe3DVMR2mlwBosoedqoL0-CGN_01CKLgMgxuw62KA,4861
 tico/interpreter/interpreter.py,sha256=tGbluCbrehTCqBu8mtGDNzby_ieJ2ry8_RH_eC0CQxk,3828
 tico/passes/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 tico/passes/cast_aten_where_arg_type.py,sha256=ybtGj1L7_2zGyfb_G-_y1N1mRgKHVq6fBZc-9-fH9sA,7229
@@ -79,7 +79,7 @@ tico/passes/fill_meta_val.py,sha256=Xbam6Aq90ZfWItZw1dgLIwH_q8RCiU5JodKNqkj-ink,
 tico/passes/fuse_leading_unsqueeze_reshape.py,sha256=88jwTP35yRyXOk9xdO6YW2OEfdKAws3KFRT16WQz0RI,4291
 tico/passes/fuse_redundant_reshape_to_mean.py,sha256=GhJS1ZKB6Ns4AhwcW3uUQ6q-0N-AzlD32B2EwusUJHg,3761
 tico/passes/legalize_causal_mask_value.py,sha256=xKdFwwMaSFCSQpSk8xISOAqFpZ1jIhgbBIqf7KTSGuk,4017
-tico/passes/legalize_predefined_layout_operators.py,sha256=6jd_FmXX5rbBxqp3H5MQoCnL3vY3qoAdXaXkVdfXEjI,15902
+tico/passes/legalize_predefined_layout_operators.py,sha256=MNx7L2dAlsxSazb-F7c0onPqHleI17zAc7AzQAa9aJ4,18934
 tico/passes/lower_pow2_to_mul.py,sha256=nfJXa9ZTZMiLg6ownSyvkM4KF2z9tZW34Q3CCWI_vmQ,2402
 tico/passes/lower_to_resize_nearest_neighbor.py,sha256=N6F56Of8Aiv-KIiYLHnh33WX72W60ZVQSBEYWHdYqNQ,9005
 tico/passes/lower_to_slice.py,sha256=0qAX3WzZdyMFDW4DiO9b5JFXd4rL1-0doBT6lJvaw_I,7260
@@ -115,11 +115,11 @@ tico/serialize/operators/op_cat.py,sha256=XDYOh0XAyrM0TlxVm6Sa0OFFGrKk7aSDcGXC-h
 tico/serialize/operators/op_clamp.py,sha256=ZRAsXLGsZqJEh4wXxESEpRJkRtUuJWTDgAem6lr9_5I,4298
 tico/serialize/operators/op_clone.py,sha256=vzDYJ8TS3tc2BAyd_z8nt5VqT1inpymSseMEhd9dva0,2394
 tico/serialize/operators/op_constant_pad_nd.py,sha256=OpP4AP-d1IFcWZolNa-o9ZxzXJQkMdG9WQ66soX3s-E,2675
-tico/serialize/operators/op_conv2d.py,sha256=BmSCunhziD9EhXEkWwFrWkaQ_t3cIhrJJQSRLbgqmxI,7338
+tico/serialize/operators/op_conv2d.py,sha256=UfYk5xnA9PqVYyjU9dUCSW0CiCmcEK3LnlnFh0WY4Gg,6599
 tico/serialize/operators/op_copy.py,sha256=vaianLQ19-2ZQZ-MdQ07YuOPeFeo_HAx2a0Qfn7I5Kk,6122
 tico/serialize/operators/op_cos.py,sha256=N12bNyuTQIxRnD0eHRPdFVzRQPMy1NFM4iM8oQ4lYzw,2034
 tico/serialize/operators/op_cumsum.py,sha256=3fmOf1mIeCX1uhTBcSJmRGXejzLtO8UwaI1eEQDC6nA,3798
-tico/serialize/operators/op_depthwise_conv2d.py,sha256=PTos0tQoM8EZoB88s4Tjb7n6pJja5nbNQRDsucVzRwc,7532
+tico/serialize/operators/op_depthwise_conv2d.py,sha256=wH1SFjhWJdJrb8xi2qCiCeSWNxlL8IjEwALGCxTQxbc,7034
 tico/serialize/operators/op_dequantize_per_channel.py,sha256=aPcVxjdgvfSFoLnv9NL-RxO5vZYj8ulqriMP5LHIWs0,3133
 tico/serialize/operators/op_dequantize_per_tensor.py,sha256=u9aK_Xle9rDN0EHLE0YrCTlXY4Q53Ch9Di4qmx7ynps,2304
 tico/serialize/operators/op_div.py,sha256=WjeM2Ux7TyGlSNx2aVC783JvcL0xnY6FBYo1Q_kdb5Q,2201
@@ -175,32 +175,34 @@ tico/serialize/operators/op_sub.py,sha256=yZskQJF0ylXVk02Uid8djPNIWDJ-0uHJar4UYh
 tico/serialize/operators/op_sum.py,sha256=B5aSwQMhyoBe2JYdE5nVQ3QeVDSzL-yuZZujsG08OdQ,2294
 tico/serialize/operators/op_tanh.py,sha256=rs7FsbQeUQ7Ak8RoQV9ymNGXHXRObojfY_SiqJiyqdA,1846
 tico/serialize/operators/op_to_copy.py,sha256=a8T0uPMavMO_md1a-4_0dlvDHyZS_xew0qB6xjf69rI,3934
+tico/serialize/operators/op_transpose_conv.py,sha256=YDObXXaHNOD7yjO1ccaB_NCfc5-L76ClvT3pduL8E90,5631
 tico/serialize/operators/op_unsqueeze.py,sha256=ZHhfVXSWEiwb2VDYX5uhxbGQyzZjKT7CrbBpVGxVHBU,2310
 tico/serialize/operators/op_view.py,sha256=5EMww-ve17Vm9XPuV03Tn7vJsjpU2J8U4d_FOrlm9_o,2546
 tico/serialize/operators/op_where.py,sha256=doE81GSwygrPBm3JIfN9w7kKXxeIYKxgk0eoY22QIcg,2845
 tico/serialize/operators/utils.py,sha256=lXGpEJW1h8U_-gfc6EWjvvSiq3yJ9P-v1v3EMRT_pSk,2954
 tico/utils/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
-tico/utils/convert.py,sha256=crjCVDLoGBSSpI0EAq-cr_oL68CcntOBy361XVcaPzU,12444
+tico/utils/convert.py,sha256=5C8Z2ia2XN4k3XgtJrFZYJSEejoeMllyr8YW6gwu9mw,12763
 tico/utils/define.py,sha256=Ypgp7YffM4pgPl4Zh6TmogSn1OxGBMRw_e09qYGflZk,1467
 tico/utils/diff_graph.py,sha256=_eDGGPDPYQD4b--MXX0DLoVgSt_wLfNPt47UlolLLR4,5272
 tico/utils/errors.py,sha256=f3csJjgbXG9W1aHhqEcou008Aor19W57X8oT5Hx8w1M,954
 tico/utils/graph.py,sha256=Y6aODsnc_-9l61oanknb7K1jqJ8B35iPypOKkM0Qkk0,9149
+tico/utils/installed_packages.py,sha256=J0FTwnkCGs0MxRWoCMYAqiwH7Z0GWFDLV--x-IndSp4,1017
 tico/utils/logging.py,sha256=IlbBWscsaHidI0dNqro1HEXAbIcbkR3BD5ukLy2m95k,1286
 tico/utils/model.py,sha256=Uqc92AnJXQ2pbvctS2z2F3Ku3yNrwXZ9O33hZVis7is,1250
-tico/utils/padding.py,sha256=jNMX2KFoZ3c6HTlMU8BAwG3Fyrqpq4F3ytKP13Pg4ps,1498
+tico/utils/padding.py,sha256=qk6eh295msjDl3yFRg2OAjF-p6q0B8U2u_rnB2odT2A,3024
 tico/utils/passes.py,sha256=kGmDe__5cPaO6i5EDAoXSVe6yXEoX9hAny4ROb3ZEmQ,2409
-tico/utils/register_custom_op.py,sha256=qheG1WqtkUaG1SnHrrKQ7-fE4IZRETApCsfMkjDKcfs,23240
+tico/utils/register_custom_op.py,sha256=3-Yl6iYmx1qQA2igNHt4hYhQhQMkdPb7gF50LIY8yvc,27350
 tico/utils/serialize.py,sha256=AQXMBOLu-Kg2Rn-qbqsAtHndjZAZIavlKA0QFgJREHM,1420
 tico/utils/trace_decorators.py,sha256=ddLIiKQfSaQrxgF1kNpwjFTQnXENzeSfcr1kuAW4jGI,3221
 tico/utils/utils.py,sha256=fnbZ2RLH6-J-wqb32O4qsR1ce4BJU0wYNrk84QXa6_E,13158
-tico/utils/validate_args_kwargs.py,sha256=ifzO4ikubDPU2iXRBPF8KeyubW23cjxBThOslLAcTrg,25368
+tico/utils/validate_args_kwargs.py,sha256=cJAK6aqdzK3_Xccu6K1FQ32WGdmwWA_SqJ--TPavIuk,26614
 tico/utils/mx/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 tico/utils/mx/elemwise_ops.py,sha256=V6glyAHsVR1joqpsgnNytatCD_ew92xNWZ19UFDoMTA,10281
 tico/utils/mx/formats.py,sha256=uzNWyu-1onUlwQfX5cZ6fZSUfHMRqorper7_T1k3jfk,3404
 tico/utils/mx/mx_ops.py,sha256=RcfUTYVi-wilGB2sC35OeARdwDqnixv7dG5iyZ-fQT8,8555
-tico-0.1.0.dev250630.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
-tico-0.1.0.dev250630.dist-info/METADATA,sha256=SM3Z2qgcIkj7qUL2DKzlp4F47pwbr-3dR1ZTL-gtdMc,8846
-tico-0.1.0.dev250630.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
-tico-0.1.0.dev250630.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
-tico-0.1.0.dev250630.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
-tico-0.1.0.dev250630.dist-info/RECORD,,
+tico-0.1.0.dev250702.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
+tico-0.1.0.dev250702.dist-info/METADATA,sha256=y6E2HN1pWGO-n-xqv9GGXkNHhNbvY7d9iIA-lWVlaZE,8846
+tico-0.1.0.dev250702.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
+tico-0.1.0.dev250702.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
+tico-0.1.0.dev250702.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
+tico-0.1.0.dev250702.dist-info/RECORD,,