tico 0.1.0.dev250917__py3-none-any.whl → 0.1.0.dev250921__py3-none-any.whl

tico/passes/convert_matmul_to_linear.py

@@ -0,0 +1,200 @@
+# 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 List, Optional, TYPE_CHECKING
+
+if TYPE_CHECKING:
+    import torch.fx
+import torch
+from torch._export.utils import is_buffer, is_lifted_tensor_constant, is_param
+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
+from tico.utils.validate_args_kwargs import MatmulArgs
+
+
+class Converter:  # type: ignore[empty-body]
+    def __init__(self):
+        super().__init__()
+
+    def match(self, exported_program, node) -> bool:  # type: ignore[empty-body]
+        return False
+
+    def convert(self, exported_program, node) -> torch.fx.Node:  # type: ignore[empty-body]
+        pass
+
+
+class MatmulToLinearConverter(Converter):
+    def __init__(self):
+        super().__init__()
+
+    def convert(self, exported_program, node) -> torch.fx.Node:
+        graph_module = exported_program.graph_module
+        graph = graph_module.graph
+
+        mm_args = MatmulArgs(*node.args, **node.kwargs)  # type: ignore[arg-type]
+
+        lhs = mm_args.input
+        rhs = mm_args.other
+
+        with graph.inserting_before(node):
+            transpose_node = create_node(
+                graph,
+                torch.ops.aten.permute.default,
+                args=(rhs, [1, 0]),
+            )
+            fc_node = create_node(
+                graph,
+                torch.ops.aten.linear.default,
+                args=(lhs, transpose_node),
+            )
+            node.replace_all_uses_with(fc_node, propagate_meta=True)
+
+        return fc_node
+
+
+class RhsConstMatmulToLinearConverter(MatmulToLinearConverter):
+    def __init__(self):
+        super().__init__()
+
+    def match(self, exported_program, node) -> bool:
+        if not node.target == torch.ops.aten.mm.default:
+            return False
+
+        mm_args = MatmulArgs(*node.args, **node.kwargs)  # type: ignore[arg-type]
+
+        rhs = mm_args.other
+        if isinstance(rhs, torch.fx.Node):
+            if is_lifted_tensor_constant(exported_program, rhs):
+                return True
+            elif is_param(exported_program, rhs):
+                return True
+            elif is_buffer(exported_program, rhs):
+                return True
+            else:
+                return False
+        return False
+
+    def convert(self, exported_program, node) -> torch.fx.Node:
+        return super().convert(exported_program, node)
+
+
+class LhsConstMatmulToLinearConverter(MatmulToLinearConverter):
+    def __init__(self):
+        super().__init__()
+
+    def match(self, exported_program, node) -> bool:
+        if not node.target == torch.ops.aten.mm.default:
+            return False
+
+        mm_args = MatmulArgs(*node.args, **node.kwargs)
+        lhs = mm_args.input
+        if isinstance(lhs, torch.fx.Node):
+            if is_lifted_tensor_constant(exported_program, lhs):
+                return True
+            elif is_param(exported_program, lhs):
+                return True
+            elif is_buffer(exported_program, lhs):
+                return True
+            else:
+                return False
+        return False
+
+    def convert(self, exported_program, node) -> torch.fx.Node:
+        return super().convert(exported_program, node)
+
+
+@trace_graph_diff_on_pass
+class ConvertMatmulToLinear(PassBase):
+    """
+    This pass converts matmul to linear selectively
+
+    How to select between `matmul` and `linear`?
+
+    * Linear has better quantization accuracy (NPU backend)
+        Due to ONE compiler's quantization policy;
+        FullyConnected(=Linear) uses per-channel quantization for weight and per-tensor for input.
+        BatchMatmul(=matmul) uses per-tensor quantization for both rhs and lhs.
+
+    * Matmul to Linear requires Transpose, which may harm latency
+        When RHS is constant, addtional transpose can be folded.
+
+    [RHS non-const case]
+    Constant folding cannot be performed.
+
+    lhs         rhs (non-const)
+    |           |
+    |           transpose
+    |           |
+     -- linear --
+         |
+         out
+
+    [RHS const case]
+    Constant folding can be performed to
+
+    lhs         rhs (const)         lh          rhs (folded const)
+    |           |                   |           |
+    |           transpose           |           |
+    |           |                   |           |
+     -- linear --         -->        -- linear --
+         |                                |
+         out                              out
+
+
+    enable_lhs_const: If true, convert matmul where LHS is constant tensor. Default is False.
+    enable_rhs_const: If true, convert matmul where RHS is constant tensor. Default is True.
+    """
+
+    def __init__(
+        self,
+        enable_lhs_const: Optional[bool] = False,
+        enable_rhs_const: Optional[bool] = True,
+    ):
+        super().__init__()
+        self.converters: List[Converter] = []
+        if enable_lhs_const:
+            self.converters.append(LhsConstMatmulToLinearConverter())
+        if enable_rhs_const:
+            self.converters.append(RhsConstMatmulToLinearConverter())
+
+    def call(self, exported_program: ExportedProgram) -> PassResult:
+        logger = logging.getLogger(__name__)
+
+        graph_module = exported_program.graph_module
+        graph = graph_module.graph
+        modified = False
+        for node in graph.nodes:
+            if not node.op == "call_function":
+                continue
+
+            for converter in self.converters:
+                if not converter.match(exported_program, node):
+                    continue
+
+                new_node = converter.convert(exported_program, node)
+                modified = True
+                logger.debug(
+                    f"{node.name} is replaced with {new_node.name} operator (permute + linear)"
+                )
+                continue
+
+        graph.eliminate_dead_code()
+        graph.lint()
+        graph_module.recompile()
+
+        return PassResult(modified)

tico/passes/convert_to_relu6.py

@@ -172,7 +172,7 @@ class ConvertToReLU6(PassBase):
                 converter.convert(exported_program, node)
                 modified = True
                 logger.debug(f"{node.name} is replaced with ReLU6 operator")
-                break
+                continue
 
         graph.eliminate_dead_code()
         graph.lint()

tico/serialize/circle_serializer.py

@@ -20,6 +20,7 @@ import torch
 from circle_schema import circle
 from torch.export.exported_program import ConstantArgument, ExportedProgram, InputKind
 
+from tico.config import CompileConfigBase, get_default_config
 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
@@ -47,7 +48,9 @@ def _initialize_model() -> tuple[CircleModel, CircleSubgraph]:
     return model, graph
 
 
-def build_circle(ep: ExportedProgram) -> bytes:
+def build_circle(
+    ep: ExportedProgram, config: CompileConfigBase = get_default_config()
+) -> bytes:
     """Convert ExportedProgram to Circle format.
 
     Args:
@@ -68,9 +71,13 @@ def build_circle(ep: ExportedProgram) -> bytes:
     for in_spec in ep.graph_signature.input_specs:
         if in_spec.kind != InputKind.USER_INPUT:
             continue
-        # NoneType ConstantArgument is ignored.
-        if isinstance(in_spec.arg, ConstantArgument) and in_spec.arg.value == None:
-            continue
+        if isinstance(in_spec.arg, ConstantArgument):
+            # ConstantArgument is ignored when option is given
+            if config.get("remove_constant_input"):
+                continue
+            # NoneType ConstantArgument is ignored.
+            if in_spec.arg.value == None:
+                continue
         arg_name = in_spec.arg.name
         graph.add_input(arg_name)
         logger.debug(f"Registered input: {arg_name}")

tico/serialize/operators/op_mm.py

@@ -20,7 +20,7 @@ if TYPE_CHECKING:
 import torch
 from circle_schema import circle
 
-from tico.serialize.circle_graph import CircleSubgraph, is_const
+from tico.serialize.circle_graph import CircleSubgraph
 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
@@ -28,9 +28,9 @@ from tico.utils.validate_args_kwargs import MatmulArgs
 
 
 @register_node_visitor
-class MatmulDefaultVisitor(NodeVisitor):
+class MatmulVisitor(NodeVisitor):
     """
-    Convert matmul to equavalent BatchMatMul or FullyConnected with Transpose.
+    Convert matmul to Circle BatchMatMul
     """
 
     target: List[torch._ops.OpOverload] = [torch.ops.aten.mm.default]
@@ -38,131 +38,7 @@ class MatmulDefaultVisitor(NodeVisitor):
     def __init__(self, op_codes: Dict[OpCode, int], graph: CircleSubgraph):
         super().__init__(op_codes, graph)
 
-    # NOTE: Matmul is equivalent to Batch MatMul (batch=1)
-    def define_bmm_node(self, inputs, outputs) -> circle.Operator.OperatorT:
-        def set_bmm_option(operator):
-            operator.builtinOptionsType = (
-                circle.BuiltinOptions.BuiltinOptions.BatchMatMulOptions
-            )
-            option = circle.BatchMatMulOptions.BatchMatMulOptionsT()
-            option.adjointLhs, option.adjointRhs = False, False
-            option.asymmetricQuantizeInputs = False
-            operator.builtinOptions = option
-
-        op_index = get_op_index(
-            circle.BuiltinOperator.BuiltinOperator.BATCH_MATMUL, self._op_codes
-        )
-        operator = create_builtin_operator(self.graph, op_index, inputs, outputs)
-        set_bmm_option(operator)
-
-        return operator
-
-    def define_transpose_node(self, inputs, outputs) -> circle.Operator.OperatorT:
-        def set_transpose_option(operator):
-            operator.builtinOptionsType = (
-                circle.BuiltinOptions.BuiltinOptions.TransposeOptions
-            )
-            option = circle.TransposeOptions.TransposeOptionsT()
-            operator.builtinOptions = option
-
-        transpose_op_index = get_op_index(
-            circle.BuiltinOperator.BuiltinOperator.TRANSPOSE, self._op_codes
-        )
-        operator = create_builtin_operator(
-            self.graph, transpose_op_index, inputs, outputs
-        )
-        set_transpose_option(operator)
-        return operator
-
-    def define_fc_node(self, inputs, outputs) -> circle.Operator.OperatorT:
-        def set_fc_option(operator):
-            operator.builtinOptionsType = (
-                circle.BuiltinOptions.BuiltinOptions.FullyConnectedOptions
-            )
-            option = circle.FullyConnectedOptions.FullyConnectedOptionsT()
-
-            option.fusedActivationFunction = (
-                circle.ActivationFunctionType.ActivationFunctionType.NONE
-            )
-            option.weightsFormat = (
-                circle.FullyConnectedOptionsWeightsFormat.FullyConnectedOptionsWeightsFormat.DEFAULT
-            )
-            option.keepNumDims = False
-            option.asymmetricQuantizeInputs = False
-            option.quantizedBiasType = circle.TensorType.TensorType.FLOAT32
-
-            operator.builtinOptions = option
-
-        fc_op_index = get_op_index(
-            circle.BuiltinOperator.BuiltinOperator.FULLY_CONNECTED, self._op_codes
-        )
-        operator = create_builtin_operator(self.graph, fc_op_index, inputs, outputs)
-        set_fc_option(operator)
-        return operator
-
-    """
-    Define FullyConnnected with Tranpose operator.
-    Note that those sets of operators are equivalent.
-    (1) Matmul
-    matmul( lhs[H, K], rhs[K, W'] ) -> output(H, W')
-    
-    (2) Transpose + FullyConneccted
-    transpose( rhs[K, W'] ) -> trs_output[W', K]
-    fullyconnected( lhs[H, K], trs_output[W', K] ) -> output(H, W')
-    """
-
-    def define_fc_with_transpose(
-        self, node, inputs, outputs
-    ) -> circle.Operator.OperatorT:
-        lhs, rhs = inputs
-
-        # get transpose shape
-        rhs_tid: int = self.graph.get_tid_registered(rhs)
-        rhs_tensor: circle.Tensor.TensorT = self.graph.tensors[rhs_tid]
-        rhs_name: str = rhs.name
-        rhs_type: int = rhs_tensor.type
-        rhs_shape: List[int] = rhs_tensor.shape
-        assert len(rhs_shape) == 2, len(rhs_shape)
-        rhs_shape_transpose = [rhs_shape[1], rhs_shape[0]]
-
-        # create transpose output tensor
-        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,
-        )
-        trs_perm = self.graph.add_const_tensor(data=[1, 0], source_node=node)
-        trs_operator = self.define_transpose_node([rhs, trs_perm], [trs_output])
-        self.graph.add_operator(trs_operator)
-
-        # define fc node
-        fc_input = lhs
-        fc_weight = trs_output
-        fc_shape = [fc_weight.shape[0]]
-        fc_bias = self.graph.add_const_tensor(
-            data=[0.0] * fc_shape[0], source_node=node
-        )
-
-        operator = self.define_fc_node([fc_input, fc_weight, fc_bias], outputs)
-
-        return operator
-
-    def define_node(
-        self, node: torch.fx.Node, prior_latency=True
-    ) -> circle.Operator.OperatorT:
-        """
-        NOTE: Possibility of accuracy-latency trade-off
-        From ONE compiler's perspective:
-        - BMM uses per-tensor quantization for both rhs and lhs.
-        - FC uses per-channel quantization for weight and per-tensor for input.
-        Thus, FC is better in terms of accuracy.
-        FC necessarily involves an additional transpose operation to be identical with mm.
-        If transposed operand is const, it can be optimized by constant folding.
-        Thus, convert FC only if tranpose can be folded.
-        TODO set prior_latency outside
-        """
+    def define_node(self, node: torch.fx.Node) -> circle.Operator.OperatorT:
         args = MatmulArgs(*node.args, **node.kwargs)  # type: ignore[arg-type]
         input = args.input
         other = args.other
@@ -170,9 +46,16 @@ class MatmulDefaultVisitor(NodeVisitor):
         inputs = [input, other]
         outputs = [node]
 
-        if not is_const(other) and prior_latency:
-            operator = self.define_bmm_node(inputs, outputs)
-        else:
-            operator = self.define_fc_with_transpose(node, inputs, outputs)
+        op_index = get_op_index(
+            circle.BuiltinOperator.BuiltinOperator.BATCH_MATMUL, self._op_codes
+        )
+        operator = create_builtin_operator(self.graph, op_index, inputs, outputs)
+        operator.builtinOptionsType = (
+            circle.BuiltinOptions.BuiltinOptions.BatchMatMulOptions
+        )
+        option = circle.BatchMatMulOptions.BatchMatMulOptionsT()
+        option.adjointLhs, option.adjointRhs = False, False
+        option.asymmetricQuantizeInputs = False
+        operator.builtinOptions = option
 
         return operator

tico/utils/convert.py

@@ -40,6 +40,7 @@ from tico.passes.cast_mixed_type_args import CastMixedTypeArgs
 from tico.passes.const_prop_pass import ConstPropPass
 from tico.passes.convert_conv1d_to_conv2d import ConvertConv1dToConv2d
 from tico.passes.convert_layout_op_to_reshape import ConvertLayoutOpToReshape
+from tico.passes.convert_matmul_to_linear import ConvertMatmulToLinear
 from tico.passes.convert_repeat_to_expand_copy import ConvertRepeatToExpandCopy
 from tico.passes.convert_to_relu6 import ConvertToReLU6
 from tico.passes.decompose_addmm import DecomposeAddmm
@@ -249,6 +250,10 @@ def convert_exported_module_to_circle(
             ConstPropPass(),
             SegmentIndexSelectConst(),
             LegalizeCausalMaskValue(enabled=config.get("legalize_causal_mask_value")),
+            ConvertMatmulToLinear(
+                enable_lhs_const=config.get("convert_lhs_const_mm_to_fc"),
+                enable_rhs_const=config.get("convert_rhs_const_mm_to_fc"),
+            ),
             LowerToResizeNearestNeighbor(),
             LegalizePreDefinedLayoutOperators(),
             LowerPow2ToMul(),
@@ -287,7 +292,7 @@ def convert_exported_module_to_circle(
 
     check_unsupported_target(exported_program)
     check_training_ops(exported_program)
-    circle_program = build_circle(exported_program)
+    circle_program = build_circle(exported_program, config)
 
     return circle_program
 

{tico-0.1.0.dev250917.dist-info → tico-0.1.0.dev250921.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: tico
-Version: 0.1.0.dev250917
+Version: 0.1.0.dev250921
 Summary: Convert exported Torch module to circle
 Home-page: UNKNOWN
 License: UNKNOWN

{tico-0.1.0.dev250917.dist-info → tico-0.1.0.dev250921.dist-info}/RECORD

@@ -1,19 +1,18 @@
-tico/__init__.py,sha256=Da7Ln6MuWCBJXrjts6OsAslWSS79toVgPG2PITYPzE0,1883
+tico/__init__.py,sha256=SJrnDNsVJlIf-r1ZVzi2Kj_xI68YVwAjm83FWgbiWLE,1883
 tico/pt2_to_circle.py,sha256=gu3MD4Iqc0zMZcCZ2IT8oGbyj21CTSbT3Rgd9s2B_9A,2767
 tico/config/__init__.py,sha256=xZzCXjZ84qE-CsBi-dfaL05bqpQ3stKKfTXhnrJRyVs,142
 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/config/v1.py,sha256=AVgOck-HxR1R1FZPVjtN5J82hPLJvUxwzbnyWXIQZWE,1237
 tico/experimental/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 tico/experimental/quantization/__init__.py,sha256=IaJPZegVJp0P3luutBo907Kp5sOJensE1Mm-XBG_jBs,122
-tico/experimental/quantization/config.py,sha256=nMepa_H471t7f3bKMvR8cZUZgruy_8kdb147rBkTWCQ,2004
-tico/experimental/quantization/public_interface.py,sha256=4-v9VXsokRG2-UUYYHd_MlbHxChqdGI5iuySyYDY_Pw,4420
-tico/experimental/quantization/quantizer.py,sha256=_2pDtWFKDCuKfYF2bptOwIYsa0VFNFM1ZNgi8_OGvHM,2365
+tico/experimental/quantization/public_interface.py,sha256=y-iwaeuedBvHwTh5hflQg4u2ZCdqf46IlTl9ntHq8pU,4425
+tico/experimental/quantization/quantizer.py,sha256=pDTQGzR-BcQJeGZ7O4cXRQdCme4q_POpxHetwnv0bYg,2370
 tico/experimental/quantization/algorithm/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 tico/experimental/quantization/algorithm/gptq/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 tico/experimental/quantization/algorithm/gptq/gptq.py,sha256=Qn9b_2ki7B64DcVEY25NMkww3PdZ5EqYQQXfYhNDQ6I,5555
 tico/experimental/quantization/algorithm/gptq/quant.py,sha256=Rl4wAOCmlE0U09BtNCDbccaSNohRHCNLwFi3zCqZfNo,5127
-tico/experimental/quantization/algorithm/gptq/quantizer.py,sha256=_ZnSD_LBag_FVcVEniPKBmw7bNZ2iZLZ8aZnexnCgrs,11693
+tico/experimental/quantization/algorithm/gptq/quantizer.py,sha256=ZKeQQWm6eMUyRgntQxVR-QVjxJOc2pW4Dc_mrEPZA64,11686
 tico/experimental/quantization/algorithm/gptq/utils.py,sha256=leGKayf-xbSjVwwAGTA5RsxUKrhDiklOQdlsLifjdrs,1811
 tico/experimental/quantization/algorithm/pt2e/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 tico/experimental/quantization/algorithm/pt2e/quantizer.py,sha256=mdTvsG87bo8fu0GaWqSM8iBCs-4f4EfUlVtk-Ko6M34,2546
@@ -38,8 +37,13 @@ tico/experimental/quantization/algorithm/pt2e/transformation/__init__.py,sha256=
 tico/experimental/quantization/algorithm/pt2e/transformation/convert_scalars_to_attrs.py,sha256=Idtoya2RcGKlgUJgC9WqNz0jH3gf6ViuPmsD9ySHbls,2253
 tico/experimental/quantization/algorithm/smoothquant/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 tico/experimental/quantization/algorithm/smoothquant/observer.py,sha256=OWBKQ3ox6PqeqgevxOjpXvb7uApoqE4YbUBelGhVSN8,3435
-tico/experimental/quantization/algorithm/smoothquant/quantizer.py,sha256=QuZBi24L-LYI26nwZd6JmTdokxr6-l_vIgZvWVdqx_o,3637
+tico/experimental/quantization/algorithm/smoothquant/quantizer.py,sha256=14-QrKAW-Rw6pIbbNaD5eORcH2fqi40-TNFGaWVakIg,3649
 tico/experimental/quantization/algorithm/smoothquant/smooth_quant.py,sha256=fxCy4m-BsSjraciSVPFlPhgsOT46RjrOgczQGb7B9TA,11561
+tico/experimental/quantization/config/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
+tico/experimental/quantization/config/base.py,sha256=xg_HCDSuMgYvMd6ENZe4Sm2SYJgMaCBj4cmqaz_lhAs,816
+tico/experimental/quantization/config/gptq.py,sha256=IUIEz5bLhsTXqoBCE1rfPec99zsRjwgpDbPW5YJqOPg,973
+tico/experimental/quantization/config/pt2e.py,sha256=9HCrraTGGZeKEN9puKV-ODi7ncV2Wjc3oe_JCO1D_Rs,850
+tico/experimental/quantization/config/smoothquant.py,sha256=fcyhu3YlOTM7fDW9lGTXh-uJOUD6CeykZj7AMCNVbak,1415
 tico/experimental/quantization/evaluation/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 tico/experimental/quantization/evaluation/backend.py,sha256=CZL9rZOA0t8cH7PHp6u9l7dGqWNvTj9bKOvwo0PVul0,692
 tico/experimental/quantization/evaluation/evaluate.py,sha256=kfa_GvFaX6DoSTAmuCImMJqF2jgqtnor5UpC7wVmGPI,7877
@@ -68,7 +72,7 @@ tico/experimental/quantization/ptq/examples/quantize_linear.py,sha256=8zq-ZJDYga
 tico/experimental/quantization/ptq/examples/quantize_llama_attn.py,sha256=cVWUSSzaZWFp5QZkNkrlpHU3kXyP84QtnZbahVml_yQ,4329
 tico/experimental/quantization/ptq/examples/quantize_llama_decoder_layer.py,sha256=mBWrjkyEovYQsPC4Rrsri6Pm1rlFmDb3NiP0DQQhFyM,5751
 tico/experimental/quantization/ptq/examples/quantize_llama_mlp.py,sha256=N1qZQgt1S-xZrdv-PW7OfXEcv0gsO2q9faOF4aD-zKo,4147
-tico/experimental/quantization/ptq/examples/quantize_with_gptq.py,sha256=w21Qao5_6SnWMuxmnZbZOoqaLQOuSnK52mHin4aedtA,6979
+tico/experimental/quantization/ptq/examples/quantize_with_gptq.py,sha256=y-SK56j4wL-9j-0jtuOqQUq4CElZtGOETp-Tg4XivUI,10438
 tico/experimental/quantization/ptq/observers/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 tico/experimental/quantization/ptq/observers/affine_base.py,sha256=e2Eba64nrxKQyE4F_WJ7WTSsk3xe6bkdGUKaoLFWGFw,4638
 tico/experimental/quantization/ptq/observers/base.py,sha256=Wons1MzpqK1mfcy-ppl-B2Dum0edXg2dWW2Lw3V18tw,3280
@@ -84,8 +88,9 @@ tico/experimental/quantization/ptq/wrappers/__init__.py,sha256=IO6FP_xYbGy0dW0HL
 tico/experimental/quantization/ptq/wrappers/ptq_wrapper.py,sha256=F9sK_DiRaXiGNHULcwIbs5EUtHz6ZJ7N4r5CWTTfhsM,2442
 tico/experimental/quantization/ptq/wrappers/quant_elementwise.py,sha256=LhEoobfvto6zKrBOKL4gmxfFFc31jHzyQV_zfps-iQM,3604
 tico/experimental/quantization/ptq/wrappers/quant_module_base.py,sha256=vkcDos_knGSS29rIZuEIWkAJLHrENbGz8nCH2-iara8,5969
-tico/experimental/quantization/ptq/wrappers/registry.py,sha256=GlVBPWPAnLRqTtemu_YOEX9WisF1eN6Mud7y1zzvpW0,5092
+tico/experimental/quantization/ptq/wrappers/registry.py,sha256=OVO5nev6J8Br9zsIX-Ut7ZgWzA9f_jk0Np9bGioXgQM,5171
 tico/experimental/quantization/ptq/wrappers/fairseq/__init__.py,sha256=Mc8FLd9DusyB_IT1vk1OYrRkngOYnYd05IvtA9ORVQc,160
+tico/experimental/quantization/ptq/wrappers/fairseq/quant_decoder_layer.py,sha256=JT79shxOhDtRFgm8jrrN6HKvyVotiytLjMjAxX-Cztg,20416
 tico/experimental/quantization/ptq/wrappers/fairseq/quant_encoder.py,sha256=r9DPUAbL2KRJ8zpMJ39Y9n6Oe79nte-mFcdjG2qEP-w,13809
 tico/experimental/quantization/ptq/wrappers/fairseq/quant_encoder_layer.py,sha256=aGr80Ku75j2H-UZ0elEa0mOQEyaAs2YJ4WJCN0lonn0,6412
 tico/experimental/quantization/ptq/wrappers/fairseq/quant_mha.py,sha256=HsigmOLeacLXc46QNeFqwQ0DwKQhNrtWTKEtLJoqXoc,15562
@@ -107,8 +112,9 @@ tico/passes/cast_mixed_type_args.py,sha256=Wd3sCDKJZwdb8GiMWKljm8X5CLFRd8eCz-dmW
 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_matmul_to_linear.py,sha256=Y_Me8YqrNumfMrB08WT4wwAoKIfKNak5y8Y10ekWe5s,6611
 tico/passes/convert_repeat_to_expand_copy.py,sha256=JbtFTmWyfJS2SSd_higP1IEhQeh7wHdN5dmTbbiFVCs,3237
-tico/passes/convert_to_relu6.py,sha256=1BJpUwUb6Zli_1y3eyJQo7dg9B1xvZ7sYjMbvEQsFJM,6442
+tico/passes/convert_to_relu6.py,sha256=9B6OLyF72tMvD-ugV7aBx6l1szwERufNBUaX34pkZ4c,6445
 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=736srs8SM8K_mLR0WG10LVMMLRkYkBM9OF0k1GCkAW0,5218
@@ -139,7 +145,7 @@ tico/passes/segment_index_select.py,sha256=VVCKNLtYRkr9n5lGnlzEuQsQ0WVxEYXGchFrD
 tico/serialize/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 tico/serialize/circle_graph.py,sha256=qvyul_HULoz7B_6RFKQ8s9RjEvMgPq-ynMVkZe8aqE4,12034
 tico/serialize/circle_mapping.py,sha256=c__AIHPi23lPugNJFolgMAKrw8j7gEeMaUQ1LAMSFnY,8542
-tico/serialize/circle_serializer.py,sha256=BGK9tltKkoL1h4rcrJUgDJIGlHst7aF3cZAKJk_GPWc,10950
+tico/serialize/circle_serializer.py,sha256=tw2xwm8tRjaFzZdaaS8Fa8Jfqz0r7Gn8L6D66m0QA0g,11228
 tico/serialize/pack.py,sha256=5HZ9kX3x6C6CyT_FWS6FRmvx_P7Dx21orjUNQxJ2xlo,1297
 tico/serialize/quant_param.py,sha256=6nbGKdqwMI9Cx9BLXJ9A9JU4qb770S8vTM1vCZRX3Eo,1342
 tico/serialize/operators/__init__.py,sha256=LIvXsNnN4yUCS2CGNQ5XW8p8oXDTV_WHWuOEAw1t6WY,990
@@ -190,7 +196,7 @@ tico/serialize/operators/op_max_pool2d_with_indices.py,sha256=i4iKZ262ytDKUt7bG9
 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=XcH15gjbP5aAl9rBKFQsVvN2GE4127zNH6_0v81_ExA,6855
+tico/serialize/operators/op_mm.py,sha256=VJJRLLYn9zAMcR2rsb86o809edyRJ7CW31waAL0ZXeI,2244
 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
@@ -228,7 +234,7 @@ tico/serialize/operators/utils.py,sha256=lXGpEJW1h8U_-gfc6EWjvvSiq3yJ9P-v1v3EMRT
 tico/serialize/operators/adapters/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 tico/serialize/operators/adapters/llama_rmsnorm.py,sha256=6t3dhfNpR03eIjsmhymF2JKd6lCf7PvInqMf77c_BOE,1139
 tico/utils/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
-tico/utils/convert.py,sha256=GgZwZtiqFzTdszfUQO0vcX39lKjs97gYwZ-Tiw_4Bbo,13222
+tico/utils/convert.py,sha256=bgk-a_gdRrrcDFFQHS_ElPdzORmfAZAgNendfzEpHOk,13501
 tico/utils/define.py,sha256=Ypgp7YffM4pgPl4Zh6TmogSn1OxGBMRw_e09qYGflZk,1467
 tico/utils/diff_graph.py,sha256=_eDGGPDPYQD4b--MXX0DLoVgSt_wLfNPt47UlolLLR4,5272
 tico/utils/dtype.py,sha256=L5Qb7qgbt0eQ5frUTvHYrRtTJb1dg4-JNEopcxCNg1U,1389
@@ -252,9 +258,9 @@ 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.dev250917.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
-tico-0.1.0.dev250917.dist-info/METADATA,sha256=WJdcwQ8suuOhdWCv9cW8_RW_qyckaOM5jEzlvi00vbM,8450
-tico-0.1.0.dev250917.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
-tico-0.1.0.dev250917.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
-tico-0.1.0.dev250917.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
-tico-0.1.0.dev250917.dist-info/RECORD,,
+tico-0.1.0.dev250921.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
+tico-0.1.0.dev250921.dist-info/METADATA,sha256=PKokhTsAtNxesEROg_vhfa6pIcl8WyFzlx-5H7RBcGk,8450
+tico-0.1.0.dev250921.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
+tico-0.1.0.dev250921.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
+tico-0.1.0.dev250921.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
+tico-0.1.0.dev250921.dist-info/RECORD,,