tico 0.1.0.dev250723__py3-none-any.whl → 0.1.0.dev250727__py3-none-any.whl

tico/__init__.py

@@ -20,8 +20,16 @@ from packaging.version import Version
 from tico.config import CompileConfigV1, get_default_config
 from tico.utils.convert import convert, convert_from_exported_program, convert_from_pt2
 
+__all__ = [
+    "CompileConfigV1",
+    "get_default_config",
+    "convert",
+    "convert_from_exported_program",
+    "convert_from_pt2",
+]
+
 # THIS LINE IS AUTOMATICALLY GENERATED BY setup.py
-__version__ = "0.1.0.dev250723"
+__version__ = "0.1.0.dev250727"
 
 MINIMUM_SUPPORTED_VERSION = "2.5.0"
 SECURE_TORCH_VERSION = "2.6.0"

tico/config/base.py

@@ -31,7 +31,7 @@ class CompileConfigBase:
         config = cls()
         for key in config_dict:
             if key in config.to_dict():
-                assert type(config.get(key)) == bool
+                assert isinstance(config.get(key), bool)
                 config.set(key, config_dict[key])
 
         return config

tico/experimental/quantization/__init__.py

@@ -1 +1,6 @@
 from tico.experimental.quantization.public_interface import convert, prepare
+
+__all__ = [
+    "convert",
+    "prepare",
+]

tico/experimental/quantization/passes/remove_weight_dequant_op.py

@@ -53,7 +53,7 @@ class ValRange:
         if isinstance(val, torch.Tensor):
             self.max = torch.max(val).item()
             self.min = torch.min(val).item()
-        elif type(val) == list:
+        elif isinstance(val, list):
             self.max = max(val)
             self.min = min(val)
         else:

tico/passes/cast_mixed_type_args.py

@@ -124,7 +124,7 @@ class CastMixedTypeArgs(PassBase):
             if rhs_val.dtype == type_to_promote:
                 ori_type = lhs_val.dtype
                 arg_to_promote = lhs
-            assert arg_to_promote != None
+            assert arg_to_promote is not None
 
             if isinstance(arg_to_promote, torch.fx.Node):
                 with graph.inserting_after(arg_to_promote):

tico/passes/decompose_addmm.py

@@ -20,7 +20,6 @@ import torch
 from torch.export import ExportedProgram
 
 from tico.serialize.circle_mapping import extract_shape
-from tico.utils import logging
 from tico.utils.graph import add_placeholder, create_node
 from tico.utils.passes import PassBase, PassResult
 from tico.utils.trace_decorators import trace_graph_diff_on_pass
@@ -59,8 +58,6 @@ class DecomposeAddmm(PassBase):
         super().__init__()
 
     def call(self, exported_program: ExportedProgram) -> PassResult:
-        logger = logging.getLogger(__name__)
-
         gm = exported_program.graph_module
         graph: torch.fx.Graph = gm.graph
         modified = False

tico/passes/decompose_fake_quantize.py

@@ -21,7 +21,6 @@ import torch
 # To import torch.ops.quantized_decomposed related operator
 from torch.export import ExportedProgram
 
-from tico.utils import logging
 from tico.utils.graph import create_node
 from tico.utils.passes import PassBase, PassResult
 from tico.utils.trace_decorators import trace_graph_diff_on_pass
@@ -65,7 +64,6 @@ class DecomposeFakeQuantize(PassBase):
         super().__init__()
 
     def call(self, exported_program: ExportedProgram) -> PassResult:
-        logger = logging.getLogger(__name__)
         modified = False
 
         gm = exported_program.graph_module

tico/passes/decompose_group_norm.py

@@ -22,7 +22,6 @@ import torch
 from torch.export import ExportedProgram
 
 from tico.serialize.circle_mapping import extract_shape
-from tico.utils import logging
 from tico.utils.graph import create_node
 from tico.utils.passes import PassBase, PassResult
 from tico.utils.trace_decorators import trace_graph_diff_on_pass
@@ -126,8 +125,6 @@ class DecomposeGroupNorm(PassBase):
         )
 
     def call(self, exported_program: ExportedProgram) -> PassResult:
-        logger = logging.getLogger(__name__)
-
         gm = exported_program.graph_module
         graph: torch.fx.Graph = gm.graph
         modified = False

tico/passes/legalize_predefined_layout_operators.py

@@ -206,7 +206,6 @@ class LegalizePreDefinedLayoutOperators(PassBase):
 
         args = ConvTranspose2DArgs(*node.args, **node.kwargs)  # type: ignore[arg-type]
         input = args.input
-        padding = args.padding
         groups = args.groups
         dilation = args.dilation
 
@@ -288,13 +287,12 @@ class LegalizePreDefinedLayoutOperators(PassBase):
         input = args.input
         weight = args.weight
         bias = args.bias
-        eps = args.eps
 
         running_mean = args.running_mean
         running_var = args.running_var
         use_input_stats = args.use_input_stats
 
-        if not (use_input_stats == True):
+        if not use_input_stats:
             raise NotYetSupportedError("Only support use_input_stats is True.")
         if not isinstance(running_mean, NoneType):
             raise NotYetSupportedError("Only support running_mean=None")
@@ -350,10 +348,6 @@ class LegalizePreDefinedLayoutOperators(PassBase):
         # max_pool2d_with_indices(Tensor self, int[2] kernel_size, int[2] stride=[], int[2] padding=0, int[2] dilation=1, bool ceil_mode=False) -> (Tensor, Tensor)
         args = MaxPool2dWithIndicesArgs(*node.args, **node.kwargs)  # type: ignore[arg-type]
         input_ = args.input
-        kernel_size = args.kernel_size
-        stride = args.stride
-        padding = args.padding
-        dilation = args.dilation
         ceil_mode = args.ceil_mode
         if ceil_mode:
             raise NotYetSupportedError("Only support non-ceil model.")
@@ -402,9 +396,6 @@ class LegalizePreDefinedLayoutOperators(PassBase):
         # avg_pool2d(Tensor self, int[2] kernel_size, int[2] stride=[], int[2] padding=0, bool ceil_mode=False, bool count_include_pad=True, int? divisor_override=None) -> (Tensor)
         args = AvgPool2dArgs(*node.args, **node.kwargs)  # type: ignore[arg-type]
         input_ = args.input
-        kernel_size = args.kernel_size
-        stride = args.stride
-        padding = args.padding
         ceil_mode = args.ceil_mode
         if ceil_mode:
             raise NotYetSupportedError("Only support non-ceil model.")

tico/passes/lower_to_resize_nearest_neighbor.py

@@ -67,7 +67,7 @@ class LowerToResizeNearestNeighbor(PassBase):
             return None
         # indices = [None, None, H index, W index]
         N, C, H, W = indices
-        if N != None or C != None:
+        if N is not None or C is not None:
             return None
         if not isinstance(H, torch.fx.Node):
             return None

tico/passes/remove_redundant_reshape.py

@@ -90,7 +90,7 @@ class RemoveRedundantReshapePattern1(PassBase):
             if len(permute.users) != 1:
                 continue
             permute_args = PermuteArgs(*permute.args, **permute.kwargs)  # type: ignore[arg-type]
-            permute_input, permute_dims = permute_args.input, permute_args.dims
+            permute_dims = permute_args.dims
             # (1xAxBxC) - `aten.permute` - (1xAxCxB)
             if permute_dims != [0, 1, 3, 2]:
                 continue
@@ -172,7 +172,7 @@ class RemoveRedundantReshapePattern2(PassBase):
             if len(permute.users) != 1:
                 continue
             permute_args = PermuteArgs(*permute.args, **permute.kwargs)  # type: ignore[arg-type]
-            permute_input, permute_dims = permute_args.input, permute_args.dims
+            permute_dims = permute_args.dims
             # (1xAxBxC) - `aten.permute` - (Bx1xAxC)
             if permute_dims != [2, 0, 1, 3]:
                 continue

tico/serialize/circle_graph.py

@@ -24,7 +24,7 @@ from torch._subclasses.fake_tensor import FakeTensor
 
 from tico.serialize.circle_mapping import (
     extract_circle_dtype,
-    extract_shape,
+    extract_circle_shape,
     str_to_circle_dtype,
     to_circle_dtype,
 )
@@ -151,15 +151,7 @@ class CircleSubgraph(circle.SubGraph.SubGraphT):
         self.name_to_node[tensor.name] = node
         assert node.meta.get("val") is not None
         tensor.type = extract_circle_dtype(node)
-        tensor.shape = list(extract_shape(node))
-
-        # Handle dynamic shape
-        if any(isinstance(s, torch.SymInt) for s in tensor.shape):
-            tensor.shapeSignature = tensor.shape.copy()
-            for idx, s in enumerate(tensor.shape):
-                if isinstance(s, torch.SymInt):
-                    tensor.shape[idx] = 1
-                    tensor.shapeSignature[idx] = -1
+        tensor.shape, tensor.shapeSignature = extract_circle_shape(node)
 
         if QPARAM_KEY in node.meta:
             tensor.quantization = to_circle_qparam(node.meta[QPARAM_KEY])
@@ -208,6 +200,7 @@ class CircleSubgraph(circle.SubGraph.SubGraphT):
         self,
         prefix: str,
         shape: List[int],
+        shape_signature: Optional[List[int]],
         dtype: int,
         qparam: Optional[QuantParam] = None,
         source_node: Optional[torch.fx.Node] = None,
@@ -230,6 +223,8 @@ class CircleSubgraph(circle.SubGraph.SubGraphT):
             A name prefix used to generate a unique tensor name.
         shape : List[int]
             The shape of the tensor.
+        shape_signature : Optional[List[int]]
+            The shape signature of the tensor to express Dynamic Shape. Defaults to `None` for Static Shape.
         dtype : int
             The Circle-compatible dtype of the tensor. Use `to_circle_dtype()` to convert.
         qparam : Optional[QuantParam]
@@ -250,14 +245,8 @@ class CircleSubgraph(circle.SubGraph.SubGraphT):
         if source_node is not None:
             self.name_to_node[tensor.name] = source_node
         tensor.shape = shape
-
-        # Handle dynamic shape
-        if any(isinstance(s, torch.SymInt) for s in tensor.shape):
-            tensor.shapeSignature = tensor.shape.copy()
-            for idx, s in enumerate(tensor.shape):
-                if isinstance(s, torch.SymInt):
-                    tensor.shape[idx] = 1
-                    tensor.shapeSignature[idx] = -1
+        if shape_signature is not None:
+            tensor.shapeSignature = shape_signature
 
         if qparam is not None:
             tensor.quantization = to_circle_qparam(qparam)
@@ -323,7 +312,7 @@ class CircleSubgraph(circle.SubGraph.SubGraphT):
         self, node: Union[torch.fx.Node, circle.Tensor.TensorT, ConstData]
     ) -> int:
         # return -1 if node is None. This is for generating CircleOutputExclude
-        if node == None:
+        if node is None:
             return -1
 
         if hasattr(node, "name") and node.name in self.name_to_tid:

tico/serialize/circle_mapping.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Tuple, TYPE_CHECKING, Union
+from typing import List, Optional, Tuple, TYPE_CHECKING, Union
 
 if TYPE_CHECKING:
     import torch.fx
@@ -128,6 +128,24 @@ def extract_shape(node: torch.fx.Node) -> torch.Size:
     return val_shape
 
 
+def extract_circle_shape(node: torch.fx.Node) -> Tuple[List[int], Optional[List[int]]]:
+    return to_circle_shape(extract_shape(node))
+
+
+def to_circle_shape(torch_shape: torch.Size) -> Tuple[List[int], Optional[List[int]]]:
+    shape: List[int] = list(torch_shape)
+    shapeSignature: Optional[List[int]] = None
+
+    if any(isinstance(s, torch.SymInt) for s in shape):
+        shapeSignature = shape.copy()
+        for idx, s in enumerate(shape):
+            if isinstance(s, torch.SymInt):
+                shape[idx] = 1
+                shapeSignature[idx] = -1
+
+    return shape, shapeSignature
+
+
 # Return stride of node
 def extract_stride(node: torch.fx.Node) -> Tuple[int, ...]:
     assert node.meta is not None
@@ -157,7 +175,8 @@ def check_if_i32_range(axis: Union[list, int]):
     return all(INT32_MIN <= val <= INT32_MAX for val in values)
 
 
-def circle_legalize_dtype_to(values, *, dtype: torch.dtype):
+# TODO: Revisit this dtype legalization function as it breaks SRP
+def circle_legalize_dtype_to(values, *, dtype: torch.dtype) -> torch.Tensor:
     """
     Legalize data types from `torch.int64` to `torch.int32`.
 

tico/serialize/circle_serializer.py

@@ -20,7 +20,7 @@ import torch
 from circle_schema import circle
 from torch.export.exported_program import ConstantArgument, ExportedProgram, InputKind
 
-from tico.serialize.circle_mapping import to_circle_dtype
+from tico.serialize.circle_mapping import to_circle_dtype, to_circle_shape
 from tico.serialize.operators import *
 from tico.serialize.circle_graph import CircleModel, CircleSubgraph
 from tico.serialize.operators.hashable_opcode import OpCode
@@ -322,9 +322,12 @@ def _handle_get_attr_node(
     if not isinstance(attr_tensor, torch.Tensor):
         raise ValueError(f"Attribute {node.target} is not a tensor")
 
+    attr_shape, attr_shape_signature = to_circle_shape(attr_tensor.shape)
+
     graph.add_tensor_from_scratch(
         prefix=node.name,
-        shape=list(attr_tensor.shape),
+        shape=attr_shape,
+        shape_signature=attr_shape_signature,
         dtype=to_circle_dtype(attr_tensor.dtype),
         source_node=node,
     )

tico/serialize/operators/op_any.py

@@ -22,6 +22,7 @@ from circle_schema import circle
 from tico.serialize.circle_graph import CircleSubgraph
 from tico.serialize.circle_mapping import (
     circle_legalize_dtype_to,
+    extract_circle_shape,
     extract_shape,
     extract_torch_dtype,
 )
@@ -99,19 +100,10 @@ class AnyVisitor(NodeVisitor):
         keepdim = args.keepdim
 
         input_shape = list(extract_shape(input))
-        output_shape = list(extract_shape(node))
-
-        dim_i32 = None
         if dim is None:
-            dims = tuple(i for i in range(0, len(input_shape)))
-            dim_i32 = tuple(
-                circle_legalize_dtype_to(dim, dtype=torch.int32) for dim in dims
-            )
-        if isinstance(dim, int):
-            dim_i32 = circle_legalize_dtype_to(dim, dtype=torch.int32)
-        if isinstance(dim, tuple):
-            dim_i32 = tuple(circle_legalize_dtype_to(d, dtype=torch.int32) for d in dim)
-        assert dim_i32 is not None
+            dim = tuple(i for i in range(0, len(input_shape)))
+
+        dim_i32 = circle_legalize_dtype_to(dim, dtype=torch.int32)
 
         inputs = [
             input,
@@ -125,9 +117,11 @@ class AnyVisitor(NodeVisitor):
         if dtype_torch in [torch.int32, torch.int64, torch.float32, torch.float64]:
             dst_dtype_circle = circle.TensorType.TensorType.BOOL
             dst_dtype_torch = torch.bool
+            dst_shape, dst_shape_signature = extract_circle_shape(input)
             ne_tensor: circle.Tensor.TensorT = self.graph.add_tensor_from_scratch(
                 prefix=f"{input.name}_ne",
-                shape=input_shape,
+                shape=dst_shape,
+                shape_signature=dst_shape_signature,
                 dtype=dst_dtype_circle,
                 source_node=input,
             )

tico/serialize/operators/op_avg_pool2d.py

@@ -22,7 +22,7 @@ import torch
 from circle_schema import circle
 
 from tico.serialize.circle_graph import CircleSubgraph
-from tico.serialize.circle_mapping import extract_circle_dtype, extract_shape
+from tico.serialize.circle_mapping import extract_circle_dtype, extract_circle_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
@@ -57,7 +57,12 @@ class AvgPool2DVisitor(NodeVisitor):
             return True
 
     def has_same_padding(self, args: AvgPool2dArgs) -> bool:
-        input_shape = list(extract_shape(args.input))
+        input_shape, input_shape_signature = extract_circle_shape(args.input)
+
+        if input_shape_signature is not None:
+            # TODO: support dynamic shapes
+            raise NotImplementedError("Dynamic shape is not supported yet")
+
         kernel_size = args.kernel_size
         stride = args.stride
         assert stride
@@ -137,7 +142,11 @@ class AvgPool2DVisitor(NodeVisitor):
                 ],
                 dtype=torch.int32,
             )
-            input_shape = list(extract_shape(input))
+            input_shape, input_shape_signature = extract_circle_shape(input)
+
+            if input_shape_signature is not None:
+                raise RuntimeError("Dynamic shape is not supported yet.")
+
             input_dtype: int = extract_circle_dtype(input)
             padded_input_shape = [
                 input_shape[0],
@@ -151,6 +160,7 @@ class AvgPool2DVisitor(NodeVisitor):
             padded_input_tensor = self.graph.add_tensor_from_scratch(
                 prefix=f"{input.name}_pad_output",
                 shape=padded_input_shape,
+                shape_signature=None,
                 dtype=input_dtype,
                 source_node=node,
             )

tico/serialize/operators/op_clamp.py

@@ -23,7 +23,7 @@ from circle_schema import circle
 from tico.passes import ops
 from tico.serialize.circle_graph import CircleSubgraph
 
-from tico.serialize.circle_mapping import extract_circle_dtype, extract_shape
+from tico.serialize.circle_mapping import extract_circle_dtype, extract_circle_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
@@ -101,12 +101,13 @@ class ClampVisitor(NodeVisitor):
             return self.define_minimum_node([input, max_val], [node])
 
         elif min_val is not None and max_val is not None:
-            input_shape = extract_shape(input)
+            input_shape, input_shape_signature = extract_circle_shape(input)
             input_dtype = extract_circle_dtype(input)
             minimum_tensor = self.graph.add_tensor_from_scratch(
                 prefix=f"{input.name}_min",
                 dtype=input_dtype,
-                shape=list(input_shape),
+                shape=input_shape,
+                shape_signature=input_shape_signature,
                 source_node=node,
             )
             minimum_opertor = self.define_minimum_node(

tico/serialize/operators/op_conv2d.py

@@ -20,7 +20,7 @@ if TYPE_CHECKING:
 import torch
 from circle_schema import circle
 
-from tico.serialize.circle_mapping import extract_circle_dtype, extract_shape
+from tico.serialize.circle_mapping import extract_circle_dtype, extract_circle_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
@@ -111,13 +111,17 @@ class Conv2dVisitor(NodeVisitor):
 
         assert groups == 1, "Only support group 1 conv2d"
 
-        input_shape = list(extract_shape(input_))
-        output_shape = list(extract_shape(node))
-        weight_shape = list(extract_shape(weight))
+        input_shape, input_shape_signature = extract_circle_shape(input_)
+        output_shape, _ = extract_circle_shape(node)
+        weight_shape, _ = extract_circle_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)
 
+        if input_shape_signature is not None:
+            # TODO: support dynamic shapes
+            raise NotImplementedError("Dynamic shape is not supported yet")
+
         pad_decision = identify_padding(padding, input_shape, output_shape, stride)
 
         conv_input: torch.fx.Node | circle.Tensor.TensorT = input_
@@ -143,6 +147,7 @@ class Conv2dVisitor(NodeVisitor):
             pad_output = self.graph.add_tensor_from_scratch(
                 prefix=f"{node.name}_input_pad_output",
                 shape=pad_output_shape,
+                shape_signature=None,
                 dtype=extract_circle_dtype(input_),
                 qparam=input_qparam,
                 source_node=node,

tico/serialize/operators/op_copy.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, Union
+from typing import Dict, List, Optional, TYPE_CHECKING, Union
 
 if TYPE_CHECKING:
     import torch._ops
@@ -52,7 +52,15 @@ class CopyVisitor(NodeVisitor):
     def __init__(self, op_codes: Dict[OpCode, int], graph: CircleSubgraph):
         super().__init__(op_codes, graph)
 
-    def check_to_do_broadcast(self, dst: List[int], src: List[int]) -> bool:
+    def check_to_do_broadcast(
+        self,
+        dst: List[int],
+        dst_sig: Optional[List[int]],
+        src: List[int],
+        src_sig: Optional[List[int]],
+    ) -> bool:
+        assert dst_sig is None
+        assert src_sig is None
         return dst != src
 
     def define_broadcast_to_node(
@@ -102,6 +110,12 @@ class CopyVisitor(NodeVisitor):
         # To connect 'dst' to Reshape node in the graph, 'dst' must be converted to Shape op.
         dst_tensor: circle.Tensor.TensorT = self.graph.get_tensor(dst)
         dst_shape: List[int] = dst_tensor.shape
+        dst_shape_signature: List[int] = dst_tensor.shapeSignature
+
+        if dst_shape_signature is not None:
+            # TODO: support dynamic shape
+            raise NotYetSupportedError("Dynamic shape is not supported yet.")
+
         dst_shape_tensor = torch.as_tensor(dst_shape, dtype=torch.int32)
 
         dst_shape_shape = [len(dst_shape)]
@@ -110,6 +124,7 @@ class CopyVisitor(NodeVisitor):
         shape_output = self.graph.add_tensor_from_scratch(
             prefix=f"{dst_name}_shape_output",
             shape=dst_shape_shape,
+            shape_signature=None,
             dtype=circle.TensorType.TensorType.INT32,
             source_node=node,
         )
@@ -119,9 +134,16 @@ class CopyVisitor(NodeVisitor):
 
         src_tensor: circle.Tensor.TensorT = self.graph.get_tensor(src)
         src_shape: List[int] = src_tensor.shape
+        src_shape_signature: List[int] = src_tensor.shapeSignature
+
+        if src_shape_signature is not None:
+            # TODO: support dynamic shape
+            raise NotYetSupportedError("Dynamic shape is not supported yet.")
 
         # The src tensor must be broadcastable with the dst tensor.
-        do_broadcast = self.check_to_do_broadcast(dst_shape, src_shape)
+        do_broadcast = self.check_to_do_broadcast(
+            dst_shape, dst_shape_signature, src_shape, src_shape_signature
+        )
         if do_broadcast:
             # create braodcastTo output tensor
             src_name: str = src.name
@@ -131,6 +153,7 @@ class CopyVisitor(NodeVisitor):
                 self.graph.add_tensor_from_scratch(
                     prefix=f"{src_name}_broadcast_to_output",
                     shape=dst_shape,
+                    shape_signature=dst_shape_signature,
                     dtype=src_type,
                     source_node=node,
                 )

tico/serialize/operators/op_cumsum.py

@@ -57,6 +57,7 @@ class CumsumVisitor(NodeVisitor):
         if input_dtype == torch.int32:
             input_tensor: circle.Tensor.TensorT = self.graph.get_tensor(input)
             input_shape: List[int] = input_tensor.shape
+            input_shape_signature: List[int] = input_tensor.shapeSignature
             cast_op_index = get_op_index(
                 circle.BuiltinOperator.BuiltinOperator.CAST, self._op_codes
             )
@@ -66,6 +67,7 @@ class CumsumVisitor(NodeVisitor):
                 prefix=cast_name,
                 dtype=cast_dtype,
                 shape=input_shape,
+                shape_signature=input_shape_signature,
                 source_node=node,
             )
             cast_operator = create_builtin_operator(

tico/serialize/operators/op_depthwise_conv2d.py

@@ -20,7 +20,7 @@ if TYPE_CHECKING:
 import torch
 from circle_schema import circle
 
-from tico.serialize.circle_mapping import extract_circle_dtype, extract_shape
+from tico.serialize.circle_mapping import extract_circle_dtype, extract_circle_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
@@ -115,12 +115,17 @@ class DepthwiseConv2dVisitor(NodeVisitor):
         dilation = args.dilation
         groups = args.groups
 
-        input_shape = list(extract_shape(input_))  # OHWI
-        output_shape = list(extract_shape(node))  # OHWI
-        weight_shape = list(extract_shape(weight))  # 1HWO
+        input_shape, input_shape_signature = extract_circle_shape(input_)  # OHWI
+        output_shape, _ = extract_circle_shape(node)  # OHWI
+        weight_shape, _ = extract_circle_shape(weight)  # 1HWO
         assert len(input_shape) == 4, len(input_shape)
         assert len(output_shape) == 4, len(output_shape)
-        assert len(weight_shape) == 4
+        assert len(weight_shape) == 4, len(weight_shape)
+
+        if input_shape_signature is not None:
+            # TODO: support dynamic shapes
+            raise NotImplementedError("Dynamic shape is not supported yet")
+
         assert weight_shape[0] == 1
         assert weight_shape[3] == output_shape[3]
         assert input_shape[3] == groups
@@ -156,6 +161,7 @@ class DepthwiseConv2dVisitor(NodeVisitor):
             pad_output = self.graph.add_tensor_from_scratch(
                 prefix=f"{node.name}_input_pad_output",
                 shape=pad_output_shape,
+                shape_signature=None,
                 dtype=extract_circle_dtype(input_),
                 qparam=input_qparam,
                 source_node=node,

tico/serialize/operators/op_index_select.py

@@ -49,7 +49,14 @@ class IndexSelectVisitor(NodeVisitor):
             self._op_codes,
         )
 
+        # TODO: Revise this to be simple
         dim_i32 = circle_legalize_dtype_to(dim, dtype=torch.int32)
+        assert (
+            dim_i32.dim() == 0 or len(dim_i32) == 1
+        ), f"dim should be scalar: {dim_i32}"
+        dim_i32_item = dim_i32.item()
+        assert isinstance(dim_i32_item, int)
+
         inputs = [input, index]
         outputs = [node]
 
@@ -57,7 +64,7 @@ class IndexSelectVisitor(NodeVisitor):
 
         operator.builtinOptionsType = circle.BuiltinOptions.BuiltinOptions.GatherOptions
         option = circle.GatherOptions.GatherOptionsT()
-        option.axis = dim_i32
+        option.axis = dim_i32_item
 
         operator.builtinOptions = option
 

tico/serialize/operators/op_instance_norm.py

@@ -73,12 +73,6 @@ class InstanceNormVisitor(NodeVisitor):
         eps = args.eps
 
         # Ignore training-related args
-        running_mean = args.running_mean
-        running_var = args.running_var
-        use_input_stats = args.use_input_stats
-        momentum = args.momentum
-        cudnn_enabled = args.cudnn_enabled
-
         input_shape = list(extract_shape(input))
         assert len(input_shape) == 4, len(input_shape)
 

tico/serialize/operators/op_log1p.py

@@ -23,7 +23,7 @@ from circle_schema import circle
 from tico.serialize.circle_graph import CircleSubgraph
 from tico.serialize.circle_mapping import (
     extract_circle_dtype,
-    extract_shape,
+    extract_circle_shape,
     extract_torch_dtype,
 )
 from tico.serialize.operators.hashable_opcode import OpCode
@@ -62,11 +62,12 @@ class Log1pVisitor(NodeVisitor):
         args = Log1pArgs(*node.args, **node.kwargs)  # type: ignore[arg-type]
         input = args.input
 
-        input_shape = list(extract_shape(input))
+        input_shape, input_shape_signature = extract_circle_shape(input)
         dst_dtype_circle = extract_circle_dtype(input)
         add_tensor: circle.Tensor.TensorT = self.graph.add_tensor_from_scratch(
             prefix=f"{input.name}_add",
             shape=input_shape,
+            shape_signature=input_shape_signature,
             dtype=dst_dtype_circle,
             source_node=node,
         )

tico/serialize/operators/op_max_pool2d_with_indices.py

@@ -22,7 +22,11 @@ import torch
 from circle_schema import circle
 
 from tico.serialize.circle_graph import CircleSubgraph
-from tico.serialize.circle_mapping import extract_circle_dtype, extract_shape
+from tico.serialize.circle_mapping import (
+    extract_circle_dtype,
+    extract_circle_shape,
+    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 (
@@ -88,7 +92,12 @@ class MaxPool2DWithIndicesVisitor(NodeVisitor):
                 ],
                 dtype=torch.int32,
             )
-            input_shape = list(extract_shape(input))
+            input_shape, input_shape_signature = extract_circle_shape(input)
+
+            if input_shape_signature is not None:
+                # TODO: support dynamic shape
+                raise NotImplementedError("Padding with dynamic shape is not supported")
+
             input_dtype: int = extract_circle_dtype(input)
             padded_input_shape = [
                 input_shape[0],
@@ -105,6 +114,7 @@ class MaxPool2DWithIndicesVisitor(NodeVisitor):
             padded_input_tensor = self.graph.add_tensor_from_scratch(
                 prefix=f"{input.name}_pad_output",
                 shape=padded_input_shape,
+                shape_signature=None,
                 dtype=input_dtype,
                 qparam=input_qparam,
                 source_node=node,

tico/serialize/operators/op_mm.py

@@ -129,6 +129,7 @@ class MatmulDefaultVisitor(NodeVisitor):
         trs_output = self.graph.add_tensor_from_scratch(
             prefix=f"{rhs_name}_transposed_output",
             shape=rhs_shape_transpose,
+            shape_signature=None,
             dtype=rhs_type,
             source_node=node,
         )

tico/serialize/operators/op_mul.py

@@ -66,10 +66,7 @@ class MulTensorVisitor(BaseMulVisitor):
         self,
         node: torch.fx.Node,
     ) -> circle.Operator.OperatorT:
-        args = MulTensorArgs(*node.args, **node.kwargs)  # type: ignore[arg-type]
-        input = args.input
-        other = args.other
-
+        _ = MulTensorArgs(*node.args, **node.kwargs)  # type: ignore[arg-type]
         operator = super().define_node(
             node,
         )
@@ -88,10 +85,7 @@ class MulScalarVisitor(BaseMulVisitor):
         self,
         node: torch.fx.Node,
     ) -> circle.Operator.OperatorT:
-        args = MulScalarArgs(*node.args, **node.kwargs)  # type: ignore[arg-type]
-        input = args.input
-        other = args.other
-
+        _ = MulScalarArgs(*node.args, **node.kwargs)  # type: ignore[arg-type]
         operator = super().define_node(
             node,
         )

tico/serialize/operators/op_pow.py

@@ -36,6 +36,7 @@ class BasePowVisitor(NodeVisitor):
         assert isinstance(node, torch.fx.Node), type(node)
         node_tensor: circle.Tensor.TensorT = self.graph.get_tensor(node)
         node_shape: List[int] = node_tensor.shape
+        node_shape_signature: List[int] = node_tensor.shapeSignature
         op_index = get_op_index(
             circle.BuiltinOperator.BuiltinOperator.CAST, self._op_codes
         )
@@ -45,6 +46,7 @@ class BasePowVisitor(NodeVisitor):
             prefix=cast_name,
             dtype=cast_dtype,
             shape=node_shape,
+            shape_signature=node_shape_signature,
             source_node=node,
         )
         cast_operator = create_builtin_operator(

tico/serialize/operators/op_repeat.py

@@ -21,7 +21,7 @@ import torch
 from circle_schema import circle
 
 from tico.serialize.circle_graph import CircleSubgraph
-from tico.serialize.circle_mapping import extract_circle_dtype, extract_shape
+from tico.serialize.circle_mapping import extract_circle_dtype, extract_circle_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
@@ -51,7 +51,10 @@ class RepeatVisitor(NodeVisitor):
             elif r < 0:
                 raise InvalidArgumentError("Only support positive repeat value")
 
-        tensor_shape = extract_shape(input)
+        tensor_shape, tensor_shape_signature = extract_circle_shape(input)
+        if tensor_shape_signature is not None:
+            # TODO: support dynamic shape
+            raise NotYetSupportedError("Repeat does not support dynamic shape yet.")
         assert len(tensor_shape) <= len(repeats)
         if len(tensor_shape) != len(repeats):
             # TODO Support len(tensor_shape) < len(repeats)
@@ -75,6 +78,7 @@ class RepeatVisitor(NodeVisitor):
                     concat_output = self.graph.add_tensor_from_scratch(
                         prefix=f"{node.name}_concat_{idx}",
                         shape=repeated_shape,
+                        shape_signature=None,  # TODO: support dynamic shape
                         dtype=tensor_dtype,
                         source_node=node,
                     )

tico/serialize/operators/op_reshape.py

@@ -66,7 +66,7 @@ class ReshapeVisitor(NodeVisitor):
             circle.BuiltinOptions.BuiltinOptions.ReshapeOptions
         )
         option = circle.ReshapeOptions.ReshapeOptionsT()
-        option.newShape = size_i32
+        option.newShape = size_i32.tolist()
 
         operator.builtinOptions = option
 

tico/serialize/operators/op_split_with_sizes.py

@@ -58,12 +58,14 @@ class SplitWithSizesVisitor(NodeVisitor):
         inputs = [input, split_sizes_i32, axis_i32]
 
         """
-        `split_with_sizes` has multiple output tensors and they are represented as `getitem`.
-        Therefore, unlike other ops, node itself doesn't become a circle tensor. Instead, each `getitem` will be
+        `split_with_sizes` has multiple output tensors along with `getitem`.
+        Unlike other ops, node itself doesn't become a circle tensor. Instead, each `getitem` will be
         a circle tensor.
-        Further, torch module having `split_with_sizes` may somtimes return selected outputs. At that time, `getitem`
-        nodes are generated only for the ouptut selected. Since one-compiler assumes that `CircleSplitV` always has 
-        all the outputs, let's add unused output tensors to compensate this restriction.
+        
+        torch module having `split_with_sizes` may return selected outputs by using `getitem`.
+        However, one-compiler assumes that `CircleSplitV` always have all outputs.
+        
+        So, let's add unused output tensors to compensate this restriction.
         """
         outputs: List[Union[circle.Tensor.TensorT, torch.fx.node.Node]] = []
         sorted_users = sorted(node.users.keys(), key=lambda x: x.args[1])  # type: ignore[arg-type, return-value]
@@ -80,11 +82,17 @@ class SplitWithSizesVisitor(NodeVisitor):
                 fake_tensor = node_val[idx]
                 assert isinstance(fake_tensor, FakeTensor)
                 shape = list(fake_tensor.size())
+
+                if any(isinstance(s, torch.SymInt) for s in shape):
+                    # TODO: support dynamic shape
+                    raise NotImplementedError("Dynamic shape is not supported yet.")
+
                 dtype = to_circle_dtype(fake_tensor.dtype)
                 tensor = self.graph.add_tensor_from_scratch(
-                    f"{node.name}_unused_{idx}",
-                    shape,
-                    dtype,
+                    prefix=f"{node.name}_unused_{idx}",
+                    shape=shape,
+                    shape_signature=None,  # TODO: support dynamic shape
+                    dtype=dtype,
                     source_node=node,
                 )
                 outputs.append(tensor)

tico/serialize/operators/op_transpose_conv.py

@@ -23,7 +23,7 @@ from circle_schema import circle
 from tico.serialize.circle_mapping import (
     circle_legalize_dtype_to,
     extract_circle_dtype,
-    extract_shape,
+    extract_circle_shape,
 )
 from tico.serialize.operators.hashable_opcode import OpCode
 from tico.serialize.operators.node_visitor import NodeVisitor, register_node_visitor
@@ -76,19 +76,21 @@ class TransposeConvVisitor(NodeVisitor):
         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))
+        input_shape, input_shape_signature = extract_circle_shape(input_)
+        output_shape, _ = extract_circle_shape(node)
+        weight_shape, _ = extract_circle_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)
 
+        if input_shape_signature is not None:
+            # TODO: support dynamic shapes
+            raise NotImplementedError("Dynamic shape is not supported yet")
+
         pad_decision = identify_padding(padding, input_shape, output_shape, stride)
 
         conv_input: torch.fx.Node | circle.Tensor.TensorT = input_
@@ -114,6 +116,7 @@ class TransposeConvVisitor(NodeVisitor):
             pad_output = self.graph.add_tensor_from_scratch(
                 prefix=f"{node.name}_input_pad_output",
                 shape=pad_output_shape,
+                shape_signature=None,
                 dtype=extract_circle_dtype(input_),
                 qparam=input_qparam,
                 source_node=node,

tico/serialize/operators/op_view.py

@@ -56,6 +56,7 @@ class ViewVisitor(NodeVisitor):
         if isinstance(size, int):
             raise Exception("scalar size conversion is not supported yet.")
 
+        # TODO: support dynamic shape
         size_i32 = circle_legalize_dtype_to(size, dtype=torch.int32)
         inputs = [input, size_i32]
         outputs = [node]
@@ -67,7 +68,7 @@ class ViewVisitor(NodeVisitor):
             circle.BuiltinOptions.BuiltinOptions.ReshapeOptions
         )
         option = circle.ReshapeOptions.ReshapeOptionsT()
-        option.newShape = size_i32
+        option.newShape = size_i32.tolist()
 
         operator.builtinOptions = option
 

tico/serialize/quant_param.py

@@ -12,6 +12,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from dataclasses import dataclass
+from typing import List, Optional
+
+import torch
+
 """
 This is a key for torch.fx.Node's meta dict to save QuantParam
 
@@ -19,11 +24,6 @@ QuantParam can be retrieved as node.meta[QPARAM_KEY]
 """
 QPARAM_KEY = "_quantization_parameters_"
 
-from dataclasses import dataclass
-from typing import List, Optional
-
-import torch
-
 
 @dataclass
 class QuantParam:

tico/utils/utils.py

@@ -130,7 +130,7 @@ def enforce_type(callable):
 
                     return True
 
-                if typing.get_origin(type_hint) == dict:
+                if typing.get_origin(type_hint) is dict:
                     if not isinstance(value, typing.get_origin(type_hint)):
                         return False
 

{tico-0.1.0.dev250723.dist-info → tico-0.1.0.dev250727.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: tico
-Version: 0.1.0.dev250723
+Version: 0.1.0.dev250727
 Summary: Convert exported Torch module to circle
 Home-page: UNKNOWN
 License: UNKNOWN

{tico-0.1.0.dev250723.dist-info → tico-0.1.0.dev250727.dist-info}/RECORD

@@ -1,11 +1,11 @@
-tico/__init__.py,sha256=e1nx3N3UdH9Ok1YI1Q0qGVr_5_AqoG4ei8NnoEsxeuk,1743
+tico/__init__.py,sha256=37Am_n1sr3CrCWSfMslOBX7P-iHXyaZ9IjQWSLozypc,1883
 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
+tico/config/base.py,sha256=q5xMqGxTUZs4mFqt5c7i_y9U00fYgdMGl9nUqIVMlCo,1248
 tico/config/factory.py,sha256=il0zqB6Lm5NX2LnG-TUhmiP9vVeZ_3TucJMorVZIodY,1324
 tico/config/v1.py,sha256=O1jzpUBDwoWpLohEpI08pJNwVB-yz3ufPrQm2_XWq4Y,1108
 tico/experimental/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
-tico/experimental/quantization/__init__.py,sha256=aBkUiNH6v6WOPXdFa1TUx8WbF3dVfPGqJNdUFYbkfSo,77
+tico/experimental/quantization/__init__.py,sha256=IaJPZegVJp0P3luutBo907Kp5sOJensE1Mm-XBG_jBs,122
 tico/experimental/quantization/config.py,sha256=h01WpP8Y-dLj6yg12pMZm3PXJqUnU2sWip5jBRc5x9Q,1604
 tico/experimental/quantization/public_interface.py,sha256=OKW8UoBMjPwiTacrWgQY9ENCh8ucPnYMSrl2R-w0pJ0,3982
 tico/experimental/quantization/quantizer.py,sha256=_2pDtWFKDCuKfYF2bptOwIYsa0VFNFM1ZNgi8_OGvHM,2365
@@ -55,24 +55,24 @@ tico/experimental/quantization/passes/insert_quantize_on_dtype_mismatch.py,sha25
 tico/experimental/quantization/passes/propagate_qparam_backward.py,sha256=TGtyW0Z2qOTgVIasBdGRgbwH31YYd6ek7OvLTmCV614,3118
 tico/experimental/quantization/passes/propagate_qparam_forward.py,sha256=RhUHGCR2RpBO5KYkQ7Z8U5u7HEwDq2wdKHLKAJCi-5c,5138
 tico/experimental/quantization/passes/quantize_bias.py,sha256=T7YxJ70N0tSK0FF9VJZA5iP0sHdnnsX9GX4AT4JDFSk,4325
-tico/experimental/quantization/passes/remove_weight_dequant_op.py,sha256=Klc_9-94tl0_AuAToKOjsWED_YPk5RB67eum0ddPX7o,6588
+tico/experimental/quantization/passes/remove_weight_dequant_op.py,sha256=gI1MtrHazWpdNfys7f1ngTTWplzluF7SA-uX0HMR5Mc,6592
 tico/interpreter/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 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=QOaet85z23ad3s9c8md5r11q9dEw4-lfJqlpM7aiBic,7228
 tico/passes/cast_clamp_mixed_type_args.py,sha256=m3_HpXLywWmWERfE5lM5PgvjBod7C4BWu_Q-TkRyO8k,5387
-tico/passes/cast_mixed_type_args.py,sha256=gsDJfqxVE2dqQnLjSEsPWWs9FGXT-tIyMzjLhxfktKA,7617
+tico/passes/cast_mixed_type_args.py,sha256=oXC0nDRnx4DJvmkzSVFVGqa4d2ICZvSarkXPgi9vTig,7621
 tico/passes/const_prop_pass.py,sha256=hDxGgJNiRjsgOArdaoeAOcOOA-nKBvA1W1zcMZQA5yg,11531
 tico/passes/convert_conv1d_to_conv2d.py,sha256=ktS3h158y9rg1sQiW8BZZbflV_dk_UdjBPQnuiOKyzg,5303
 tico/passes/convert_layout_op_to_reshape.py,sha256=sCAFjkmVtiKjvDQSAgnjNBHl3_hWXJZElGDXQiTH-7s,2963
 tico/passes/convert_repeat_to_expand_copy.py,sha256=JbtFTmWyfJS2SSd_higP1IEhQeh7wHdN5dmTbbiFVCs,3237
 tico/passes/convert_to_relu6.py,sha256=1BJpUwUb6Zli_1y3eyJQo7dg9B1xvZ7sYjMbvEQsFJM,6442
-tico/passes/decompose_addmm.py,sha256=0IUZjRS8G-h7LPeYu27gxWyTKi3pKAF1uNU1y-mHLL0,4056
+tico/passes/decompose_addmm.py,sha256=KjnpZjSuA0uvNmKaTN_EMwobcOi3CAB81buORzTDxro,3979
 tico/passes/decompose_batch_norm.py,sha256=06LAxhSmpTxFZJmUelwB3I_GipNWrLoM7PfM6ZkxOZY,6512
-tico/passes/decompose_fake_quantize.py,sha256=dvXZ5LyjrbpgvasmcBgjk2-xJyijefPRoNXBjFmAkRI,5294
+tico/passes/decompose_fake_quantize.py,sha256=736srs8SM8K_mLR0WG10LVMMLRkYkBM9OF0k1GCkAW0,5218
 tico/passes/decompose_fake_quantize_tensor_qparams.py,sha256=k9MJhMVABFNF6lXgEum1fJyGpdQwVRKxWOYhkMR2M7c,13915
-tico/passes/decompose_group_norm.py,sha256=rtjv3PrcFTtZ68uCsmE2LibnWaQU-e8NdKkblL4LqxE,10197
+tico/passes/decompose_group_norm.py,sha256=6BqvYtMTPzeIgp8cPA8OFMwEBvb7odcg04IUgwtp7NQ,10120
 tico/passes/decompose_grouped_conv2d.py,sha256=n2qv320akL1ju33ucZ6lU1cKEAaj0NI8YZ5CrUnkRLM,8512
 tico/passes/decompose_slice_scatter.py,sha256=xqMHKhW2595YoAeubKZ4jRhYW4TQ09EXPgLNgODqXG8,5653
 tico/passes/extract_dtype_kwargs.py,sha256=ObpsaFlrTPYQw2hJ7UsC5CocyAtBkT_bMtzkMUqAyKc,4333
@@ -80,9 +80,9 @@ 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=0nfUKGd7XSe9Hg5TAi4dUi6Nn6-JRTWCwhULR5AEgqs,4079
-tico/passes/legalize_predefined_layout_operators.py,sha256=MdpI47_AI6PDqxWHGlJCsjbg_qxJhDzMhTl1fWn8I04,18936
+tico/passes/legalize_predefined_layout_operators.py,sha256=3gILn38jzIMDXtMTWpjdROgwmavDC64w115W171encg,18641
 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_resize_nearest_neighbor.py,sha256=gbrvTmWSXDPdJ1XJtWGI5mo-uEiauXEG3ELwbKYVPLI,9013
 tico/passes/lower_to_slice.py,sha256=OzlFzK3lBYyYwC3WThsWd94Ob4JINIJF8UaLAtnumzU,7262
 tico/passes/merge_consecutive_cat.py,sha256=ayZNLDA1DFM7Fxxi2Dmk1CujkgUuaVCH1rhQgLrvvOQ,2701
 tico/passes/ops.py,sha256=cSj3Sk2x2cOE9b8oU5pmSa_rHr-iX2lORzu3N_UHMSQ,2967
@@ -90,37 +90,37 @@ tico/passes/remove_nop.py,sha256=Hf91p_EJAOC6DyWNthash0_UWtEcNc_M7znamQfYQ5Y,268
 tico/passes/remove_redundant_assert_nodes.py,sha256=IONd3xBy6I8tH6_Y1eN3_eCHH7WTC8soBgjXzOju9cQ,1612
 tico/passes/remove_redundant_expand.py,sha256=auyqIoQT4HJhiJfuUe6BrEtUhvz221ohnIK5EuszWeg,2112
 tico/passes/remove_redundant_permute.py,sha256=98UsaZzFZdQzEEAR1pIzRisAf6hgfXLa88aayjalt3E,4292
-tico/passes/remove_redundant_reshape.py,sha256=OnhU7FKpudGEaF6ufsMvkzf9hn84FNmNKk37M7FN3Pc,16454
+tico/passes/remove_redundant_reshape.py,sha256=aeep6LDvY58GEuOrWckkEXnJa6wkkbiJ9FrimT9F3-s,16384
 tico/passes/remove_redundant_slice.py,sha256=Iv7TbB39fktNb4eq0VdyZnwxL_VsKLJ90diMmaf3kZk,2087
 tico/passes/remove_redundant_to_copy.py,sha256=tKy4XKkO2l33fMxVPQ_iFkUeFvP15kbPvzPPhT_g0c8,3292
 tico/passes/restore_linear.py,sha256=xGJdNb-1CrkOKS9BnLbcblkZc6P2vVjKIi-7lRcs7Bk,4111
 tico/passes/segment_index_select.py,sha256=VVCKNLtYRkr9n5lGnlzEuQsQ0WVxEYXGchFrDnB1C40,5189
 tico/serialize/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
-tico/serialize/circle_graph.py,sha256=3t78g5eKzhHKvIBJqQ-CcwbqoV-2QwAdd_8wm4W1yXw,12317
-tico/serialize/circle_mapping.py,sha256=C9C3ORACQOdvBdnt5KRzlT8zao_TvzQklIxH794OhP0,5719
-tico/serialize/circle_serializer.py,sha256=eb90eTgU6WAZEWPclZIkInJsqAP1KjRSF2QeY5F3c3E,10766
+tico/serialize/circle_graph.py,sha256=GSrrv6UGs2-sH5wJ19M-JLNimC-ZDsPzAO_ZswlHEU4,11945
+tico/serialize/circle_mapping.py,sha256=FEOGme-ewmZyTl_V34ErF7CI-KnDb98xmK_OcOKl4PA,6417
+tico/serialize/circle_serializer.py,sha256=bVKxAex_Nmc4O9RUFWIB_KhAuubVbhhA19CS3x967y8,10891
 tico/serialize/pack.py,sha256=5HZ9kX3x6C6CyT_FWS6FRmvx_P7Dx21orjUNQxJ2xlo,1297
-tico/serialize/quant_param.py,sha256=s97GJyDOZULnqFUWPakHais31G_qqPuO0awPHCkZDvI,1342
+tico/serialize/quant_param.py,sha256=6nbGKdqwMI9Cx9BLXJ9A9JU4qb770S8vTM1vCZRX3Eo,1342
 tico/serialize/operators/__init__.py,sha256=LIvXsNnN4yUCS2CGNQ5XW8p8oXDTV_WHWuOEAw1t6WY,990
 tico/serialize/operators/hashable_opcode.py,sha256=sDVKNTgIQw4IBtMEjNH8tzssMPx1x8-U2oagomRjGF0,1368
 tico/serialize/operators/node_visitor.py,sha256=UYyCwXqSCeRyimThMShstHnt7vKM9tsuzQ_02uEwF9I,2356
 tico/serialize/operators/op_abs.py,sha256=Y-vy7rcqPT-qD3QS5R8zbApWWTPpjY6xuMMVnbIhYmQ,1827
 tico/serialize/operators/op_add.py,sha256=otm062DMHVAThWmOtSTZdPyP3P5-2cv5VL_UWBJeLms,2346
 tico/serialize/operators/op_alias_copy.py,sha256=Xu9OiILbGf8oddh8yTqovvLfgVs8XYV7Cg4n6CesWcg,2175
-tico/serialize/operators/op_any.py,sha256=vZ37qzSEAOVPjfakSkgPDym91S9kOMUpLisAqkP-3Po,5211
+tico/serialize/operators/op_any.py,sha256=wrTXFQ1TWl-2ET2NGXAXI1dzfDDJsYtA71pyj2numPE,4968
 tico/serialize/operators/op_arange_start_step.py,sha256=0T5lWwh3TfsFStmVv0v5qG03KENRDBmMix08RXQ4D-U,2132
 tico/serialize/operators/op_argmax.py,sha256=ARyGHlmWVmzwCct93V5x1-VyKqhxMOvV8GuM8yQWXdo,2290
-tico/serialize/operators/op_avg_pool2d.py,sha256=vc7WCakGXtGFPV1ix5EJmboH23tQ-cSI36ePY3PHKI4,7544
+tico/serialize/operators/op_avg_pool2d.py,sha256=OlxEezOtqdmc76d88Q5zziBBoOaIwtsHyhHMYKXHHog,7930
 tico/serialize/operators/op_bmm.py,sha256=AELjHC9ISFPIzEEl5Kr1s4GSNLZElwZmVZJWkEyCEoA,2189
 tico/serialize/operators/op_cat.py,sha256=XDYOh0XAyrM0TlxVm6Sa0OFFGrKk7aSDcGXC-hYX4gs,2204
-tico/serialize/operators/op_clamp.py,sha256=6iDFZZTMkEqJ8lpNcEhKUpAIc4h2oPkBCBWnnt4lTDw,4322
+tico/serialize/operators/op_clamp.py,sha256=RRQVrzayDfN3PioCVJqa_yYOtcYwb5HHwkMe4E_YPmE,4408
 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=UfYk5xnA9PqVYyjU9dUCSW0CiCmcEK3LnlnFh0WY4Gg,6599
-tico/serialize/operators/op_copy.py,sha256=vaianLQ19-2ZQZ-MdQ07YuOPeFeo_HAx2a0Qfn7I5Kk,6122
+tico/serialize/operators/op_conv2d.py,sha256=I9OgOr1cBGFkx1Q6eyLLvY5DmKdaieMSNmVAuZ8pBa0,6842
+tico/serialize/operators/op_copy.py,sha256=z0GCaRP4dyJdJ1ajQ45BKmBRJlirWUPqYfnrqBtbeUM,6921
 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=wH1SFjhWJdJrb8xi2qCiCeSWNxlL8IjEwALGCxTQxbc,7034
+tico/serialize/operators/op_cumsum.py,sha256=Y8eVvtX57uPA4NIwS-Vbfjoiz_rPpc_wmViTQtCoz6s,3928
+tico/serialize/operators/op_depthwise_conv2d.py,sha256=S2naBMWAoUL3nJiV7RVg97bUvuQdB5VldE_r--rX0hA,7297
 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
@@ -134,33 +134,33 @@ tico/serialize/operators/op_ge.py,sha256=TrgZ6wbIEYkDGfVFNtDlfM7ZkMMWjvcks5U5Dan
 tico/serialize/operators/op_gelu.py,sha256=bS8-0rg5_bT__OI3mBDywxGx4xTO2Iqea3h-uC17MpU,2145
 tico/serialize/operators/op_gt.py,sha256=JAVbtuAUNLYhtJycJJCEkYo9QAvmiK4lTMdw5yHUd10,1886
 tico/serialize/operators/op_index.py,sha256=iDW2YSeUS_kLiWEaQ_MjrYpxZAFBbm7_GU_2B4SRe6c,3033
-tico/serialize/operators/op_index_select.py,sha256=cw7IbvixooikGxzbpUmI9tHS4kjl4lXLtO9D-GO8qLQ,2277
-tico/serialize/operators/op_instance_norm.py,sha256=AhcVm71ChB16BlPNwqBh5tMHCqMShOXHPkE8Ag9jBfQ,3144
+tico/serialize/operators/op_index_select.py,sha256=O2MXXWGnCgS8QG3DrWKdYKbl88VBVscmOuoGcgBEf_0,2522
+tico/serialize/operators/op_instance_norm.py,sha256=5QvLefa74BrAPsTNYsi4Y7IB8d1wer4gtWantKo2nlQ,2940
 tico/serialize/operators/op_leaky_relu.py,sha256=UJPoL7kAIp6nAjyDdda_afdOcMLHme7NE77b2y76exc,2160
 tico/serialize/operators/op_linear.py,sha256=bw_mn2CiJy8CbpPevOV0PMPh0ZMWKAybLZ9cnIKJSsk,2527
 tico/serialize/operators/op_log.py,sha256=1TKvH2lttdAHE0P84vcxmOvGBBRUs6D71Jrei7SdZHE,1827
-tico/serialize/operators/op_log1p.py,sha256=gG7Fs4UDj_Nnp7U60UtPyz0fLv1lBpJVOGGCMm-42pY,3121
+tico/serialize/operators/op_log1p.py,sha256=c-fSBkaDFZ2Z_4LcZMnEvKCfGZbNOrWBMz6-Hdw98V8,3203
 tico/serialize/operators/op_logical_and.py,sha256=WhQ8knuq32BO-WhAqkOgpcUStPkjoPmRWuYNsKveF3w,2163
 tico/serialize/operators/op_logical_not.py,sha256=ugrVcRqR3IvUUaiRVW5cArCYJbzmkcXp88QM846jCww,2129
 tico/serialize/operators/op_lt.py,sha256=_vA7dWpV9wVBxB7JL9pLQT9BsV91NGQBq_0auAtHK5Y,2080
 tico/serialize/operators/op_max_dim.py,sha256=nS_TZl5uq4uv1LwgBD9Wddyac4atKqBiIWKIyeXse2s,2519
-tico/serialize/operators/op_max_pool2d_with_indices.py,sha256=Vab8KV4w0i70P5XPdqItXEv_hLFjscVngypOltRvBV8,5746
+tico/serialize/operators/op_max_pool2d_with_indices.py,sha256=2Q1upWdChLpq1x7IaoP0eRNgqTt7dH7yqlRjhEKlnjM,6034
 tico/serialize/operators/op_maximum.py,sha256=JjBr6gWEnuakLuk1_feotTHfIIm3s5YqWmqhUMpSPI0,1873
 tico/serialize/operators/op_mean.py,sha256=rVQZOxCJkHFY4kQBAS1HVK0HkcqxgkSy6zvEDLX_WYQ,2267
 tico/serialize/operators/op_minimum.py,sha256=fASjQVcTPCin02umQwFPdq2ss-Ve7S5A33J3QmmQ_wQ,1873
-tico/serialize/operators/op_mm.py,sha256=Fgq_HUUKuXOQY_t8lah3SOUqTsGet-KbVttCK4-fjAk,6821
-tico/serialize/operators/op_mul.py,sha256=42Guc0MWBGBCZoj9-4LcLtTMtUPwsmDSVmvkR8tqLhM,3165
+tico/serialize/operators/op_mm.py,sha256=XcH15gjbP5aAl9rBKFQsVvN2GE4127zNH6_0v81_ExA,6855
+tico/serialize/operators/op_mul.py,sha256=si_VdYNyFbULb50SnXHOINh0dZQ2PhRB6Fzl54ZBj5Y,3049
 tico/serialize/operators/op_ne.py,sha256=xa2WJL2tYksxw7fIJic_D9ltLEseyCII8HpR32Oq8Do,1900
 tico/serialize/operators/op_neg.py,sha256=fkI3ExyD3QF-qtxBcXqQutPNDbNL8g7lZYE7CyD2wLk,2046
 tico/serialize/operators/op_permute.py,sha256=5DfX3pfZ5FDNmrSqx3-hRwPA7vm36z7BfG-nuyyBTsM,2282
-tico/serialize/operators/op_pow.py,sha256=z_4G_J1k_keeVE6ZYKSy-kqkdJ_i4p4kHkO0dJZnz-Y,5434
+tico/serialize/operators/op_pow.py,sha256=G_GpdddmlE3opaCJtcfkV0qAWMau5uBa1sEB46qrNdo,5553
 tico/serialize/operators/op_prelu.py,sha256=0ZybL5pNvBrRvQGy4M6gELrjiEXEsb2wBDdU8x4D75I,1874
 tico/serialize/operators/op_quantize_per_tensor.py,sha256=w-vYxSPnN2gtx-pEkkcMGU0ZjiwaS4y1sxy56pKEq3E,3004
 tico/serialize/operators/op_reciprocal.py,sha256=6b9_bxjg_0EvgAitSv1MgBi4PJSEgm-21s5qtWI1UR4,2394
 tico/serialize/operators/op_relu.py,sha256=WXCR_chwEUBqjFIQ_4E2avwk-Acy76pmX20rJQCBTQo,1832
 tico/serialize/operators/op_relu6.py,sha256=ZWqEolfAKjOdUC1ZCg0iuu4dBhkJRxVYR2tUzpbvKQM,1829
-tico/serialize/operators/op_repeat.py,sha256=0wTv1Mg7kg0eHz0CT6atyVAli4T4h5rYXq5opY6op20,4235
-tico/serialize/operators/op_reshape.py,sha256=0_bJwimiGAHaKkfwfhxUw9Gebt5tnecGaEVoKhEvV0Q,2550
+tico/serialize/operators/op_repeat.py,sha256=q05RQOheLxkWgj63Czrp6tGpjGuB0900FNRKLxXdNJ4,4524
+tico/serialize/operators/op_reshape.py,sha256=6wErQpmDX9mAmfJRCTg_cg1uOdJZqHm8Nux8dNI53Vg,2559
 tico/serialize/operators/op_resize_nearest_neighbor.py,sha256=dXaAnZ5M_ko_tH-HolxNpHFXkDUQ8x45myskojP5XZE,2771
 tico/serialize/operators/op_round.py,sha256=pe6w_TB4xGLu0iPv4Qo0a0fIkY9DgCgXk5127TWt8pE,1837
 tico/serialize/operators/op_rsqrt.py,sha256=yl2vd8InjhLPbE0vHIrEera6DVXlY9dLgO7yZZCH3RI,1837
@@ -170,16 +170,16 @@ tico/serialize/operators/op_sigmoid.py,sha256=ZubbGG1yU5uvNkEmOmbjj3eq7d9mwEaJdC
 tico/serialize/operators/op_sin.py,sha256=MbttmHTVKhwKK6gH9Vbcbn5aAaxnQ71NdpmQAlTcojU,1827
 tico/serialize/operators/op_slice.py,sha256=g0r8lj5CIxpT6ixOKqUzwKiNhoiuIFwWjbpaiCoOg6w,5259
 tico/serialize/operators/op_softmax.py,sha256=qwYke5zfhnSL89DZbzdr5Fc9SsJf0vI-LDZjT_NFpbc,3669
-tico/serialize/operators/op_split_with_sizes.py,sha256=TgYg1cu-3BSz9SsXfAhoJbo4q5ZzFaoFArkH_obsYlU,4274
+tico/serialize/operators/op_split_with_sizes.py,sha256=DzSnMEsBoWpmun-NAW-lS-gssVUhaclzJ_nTxL5zZtM,4491
 tico/serialize/operators/op_sqrt.py,sha256=9Q5jkuEPrim11XfSQHGDGVTMYk1TQhOfVqMVYD_eIrI,1871
 tico/serialize/operators/op_squeeze.py,sha256=QnNwfAdTC1xBm04C9DkVs8VB5YRN-4fCsIWn189QaPg,2416
 tico/serialize/operators/op_sub.py,sha256=yZskQJF0ylXVk02Uid8djPNIWDJ-0uHJar4UYhlJVkk,2479
 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_transpose_conv.py,sha256=PgM0mSIeiCMk-V2jQ2L1JdF-v3mHwdsdY4nTEp2GRro,5796
 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_view.py,sha256=xxE-GvTJ1UpcHst5KXYz3qKY-eJQvXKKrSZiA2O7E40,2593
 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
@@ -199,15 +199,15 @@ tico/utils/register_custom_op.py,sha256=3-Yl6iYmx1qQA2igNHt4hYhQhQMkdPb7gF50LIY8
 tico/utils/serialize.py,sha256=cBtEUfi_SU_9_v0cq2CNikzn8GnzEz2RwRvUH2NkWu4,1378
 tico/utils/torch_compat.py,sha256=oc6PztVsXdHcQ3iaVR90wLLxrGaj6zFHWZ8K9rRS6q8,1795
 tico/utils/trace_decorators.py,sha256=ddLIiKQfSaQrxgF1kNpwjFTQnXENzeSfcr1kuAW4jGI,3221
-tico/utils/utils.py,sha256=Ln6FgrnWuwvKEL1ntzZXgH1tClcaOfQHt117dQ_mSb8,12976
+tico/utils/utils.py,sha256=A5p3iAAxRGDsZJh4ybp-Qo3MX3vk5RrmSY-R3rXqVeI,12976
 tico/utils/validate_args_kwargs.py,sha256=CRj_SXMUUn6onsl8XLAt-zPZCFxR4C0XOCoaad_ZD4I,26689
 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.dev250723.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
-tico-0.1.0.dev250723.dist-info/METADATA,sha256=MGuH7gedgNwadikvqXLQL_zTHDmzk20cwDohwgRkreg,8430
-tico-0.1.0.dev250723.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
-tico-0.1.0.dev250723.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
-tico-0.1.0.dev250723.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
-tico-0.1.0.dev250723.dist-info/RECORD,,
+tico-0.1.0.dev250727.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
+tico-0.1.0.dev250727.dist-info/METADATA,sha256=gE7opbcJdg8CiBvfdcjjvCuyrfQvy-q4BaBt8meVP84,8430
+tico-0.1.0.dev250727.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
+tico-0.1.0.dev250727.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
+tico-0.1.0.dev250727.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
+tico-0.1.0.dev250727.dist-info/RECORD,,