tico 0.1.0.dev250617__py3-none-any.whl → 0.1.0.dev250619__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.dev250617"
+__version__ = "0.1.0.dev250619"
 
 MINIMUM_SUPPORTED_VERSION = "2.5.0"
 SECURE_TORCH_VERSION = "2.6.0"

tico/experimental/quantization/passes/quantize_bias.py

@@ -0,0 +1,123 @@
+# 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 TYPE_CHECKING
+
+if TYPE_CHECKING:
+    import torch.fx
+import copy
+
+import torch
+from torch.export import ExportedProgram
+
+from tico.serialize.quant_param import QPARAM_KEY, QuantParam, to_qparam_dtype
+from tico.utils import logging
+from tico.utils.graph import add_placeholder, get_torch_param_value, is_torch_param
+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 LinearArgs
+
+
+@trace_graph_diff_on_pass
+class QuantizeBias(PassBase):
+    """
+    Quantize bias.
+
+    This pass identifies fp32 biases, quantizes them using scales of input and weights.
+
+    This pass assumes that if bias is fp32, input and weights must have been quantized.
+    """
+
+    def __init__(self):
+        super().__init__()
+
+    def call(self, exported_program: ExportedProgram) -> PassResult:
+        logger = logging.getLogger(__name__)
+
+        graph_module = exported_program.graph_module
+        graph: torch.fx.Graph = graph_module.graph
+        for node in graph.nodes:
+            if node.op != "call_function":
+                continue
+            if node.target == torch.ops.aten.linear.default:
+                lin_args = LinearArgs(*node.args, **node.kwargs)
+                inp = lin_args.input
+                weights = lin_args.weight
+                bias = lin_args.bias
+
+                if bias is None:
+                    continue
+
+                # Only support bias is Parameter
+                # TODO Is it possible that bias is not Parameter?
+                if not is_torch_param(bias, exported_program):
+                    continue
+
+                bias_val: torch.Tensor = get_torch_param_value(bias, exported_program)
+                if bias_val.dtype != torch.float32:
+                    continue
+
+                if QPARAM_KEY not in inp.meta:
+                    continue
+
+                if QPARAM_KEY not in weights.meta:
+                    continue
+
+                quant_dtype = None
+                if inp.meta[QPARAM_KEY].dtype == "int16":
+                    quant_dtype = torch.int64
+                elif inp.meta[QPARAM_KEY].dtype == "uint8":
+                    quant_dtype = torch.int32
+                else:
+                    continue
+
+                type_info = torch.iinfo(quant_dtype)
+
+                assert quant_dtype is not None
+
+                i_scale = inp.meta[QPARAM_KEY].scale
+                w_scale = weights.meta[QPARAM_KEY].scale
+
+                assert i_scale is not None
+                assert w_scale is not None
+                assert len(i_scale) == 1
+                assert len(w_scale) == bias_val.shape[0]
+
+                bias_scale = torch.tensor(i_scale) * torch.tensor(w_scale)
+                q_bias = torch.round(bias_val / bias_scale)
+                q_bias = torch.clamp(q_bias, min=type_info.min, max=type_info.max)
+                q_bias = q_bias.to(quant_dtype)
+
+                q_bias_node = add_placeholder(exported_program, q_bias, bias.name)
+
+                qparam = QuantParam()
+                qparam.scale = bias_scale.tolist()
+                assert qparam.scale is not None
+                qparam.zero_point = [0] * len(qparam.scale)
+                qparam.dtype = to_qparam_dtype(quant_dtype)
+                qparam.quantized_dimension = 0
+                q_bias_node.meta[QPARAM_KEY] = qparam
+
+                node.update_arg(2, q_bias_node)
+
+                logger.debug(f"Bias ({bias.name}) is quantized to {q_bias_node.name}.")
+
+            # TODO Support more ops.
+
+        graph.eliminate_dead_code()
+        graph.lint()
+        graph_module.recompile()
+
+        # Run only once.
+        return PassResult(False)

tico/experimental/quantization/passes/remove_weight_dequant_op.py

@@ -145,6 +145,9 @@ class RemoveWeightDequantOp(PassBase):
             if isinstance(dq_args, DequantizePerChannelArgs):
                 scales = get_constant(exported_program, dq_args.scales)
                 zero_ps = get_constant(exported_program, dq_args.zero_points)
+
+                # Sometimes users can give fp32 zero point. Let's update dtype here.
+                zero_ps = zero_ps.to(torch.int64)
                 quant_param.scale = scales.tolist()
                 quant_param.zero_point = zero_ps.tolist()
                 assert quant_param.zero_point is not None  # To avoid mypy error

tico/serialize/operators/op_leaky_relu.py

@@ -0,0 +1,60 @@
+# 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, TYPE_CHECKING
+
+if TYPE_CHECKING:
+    import torch._ops
+    import torch.fx
+import torch
+from circle_schema import circle
+
+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
+from tico.utils.validate_args_kwargs import LeakyReluArgs
+
+
+@register_node_visitor
+class LeakyReluVisitor(NodeVisitor):
+    target: List[torch._ops.OpOverload] = [torch.ops.aten.leaky_relu.default]
+
+    def __init__(self, op_codes: Dict[OpCode, int], graph: CircleSubgraph):
+        super().__init__(op_codes, graph)
+
+    def define_node(
+        self,
+        node: torch.fx.Node,
+    ) -> circle.Operator.OperatorT:
+        op_index = get_op_index(
+            circle.BuiltinOperator.BuiltinOperator.LEAKY_RELU, self._op_codes
+        )
+
+        args = LeakyReluArgs(*node.args, **node.kwargs)  # type: ignore[arg-type]
+        input = args.input
+        negative_slope = args.negative_slope
+
+        inputs = [input]
+        outputs = [node]
+
+        operator = create_builtin_operator(self.graph, op_index, inputs, outputs)
+        operator.builtinOptionsType = (
+            circle.BuiltinOptions.BuiltinOptions.LeakyReluOptions
+        )
+        option = circle.LeakyReluOptions.LeakyReluOptionsT()
+        option.alpha = negative_slope
+        operator.builtinOptions = option
+
+        return operator

tico/serialize/operators/op_max_pool2d_with_indices.py

@@ -124,7 +124,8 @@ class MaxPool2DWithIndicesVisitor(NodeVisitor):
                 padding_type = PaddingType.SAME
             else:
                 padding_type = PaddingType.VALID
-                maxpool_input = define_padding_node()
+                if padding[0] != 0 or padding[1] != 0:
+                    maxpool_input = define_padding_node()
 
         inputs = [maxpool_input]
         outputs = [node]

tico/utils/convert.py

@@ -30,6 +30,7 @@ from tico.experimental.quantization.passes.propagate_qparam_backward import (
 from tico.experimental.quantization.passes.propagate_qparam_forward import (
     PropagateQParamForward,
 )
+from tico.experimental.quantization.passes.quantize_bias import QuantizeBias
 from tico.experimental.quantization.passes.remove_weight_dequant_op import (
     RemoveWeightDequantOp,
 )
@@ -250,6 +251,7 @@ def convert_exported_module_to_circle(
                 RemoveWeightDequantOp(),
                 PropagateQParamForward(),
                 PropagateQParamBackward(),
+                QuantizeBias(),
                 InsertQuantizeOnDtypeMismatch(),
             ]
         )

tico/utils/validate_args_kwargs.py

@@ -513,6 +513,17 @@ class InstanceNormArgs:
     cudnn_enabled: bool
 
 
+@enforce_type
+@dataclass
+class LeakyReluArgs:
+    """
+    leaky_relu(Tensor self, Scalar negative_slope=0.01) -> Tensor
+    """
+
+    input: torch.fx.Node
+    negative_slope: float = 0.01
+
+
 @enforce_type
 @dataclass
 class LinearArgs:

{tico-0.1.0.dev250617.dist-info → tico-0.1.0.dev250619.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: tico
-Version: 0.1.0.dev250617
+Version: 0.1.0.dev250619
 Summary: Convert exported Torch module to circle
 Home-page: UNKNOWN
 License: UNKNOWN

{tico-0.1.0.dev250617.dist-info → tico-0.1.0.dev250619.dist-info}/RECORD

@@ -1,4 +1,4 @@
-tico/__init__.py,sha256=hljp9QJt0lsZ8rIFSeiTKhdyulprC_7xMjqxxA5qJ8U,1743
+tico/__init__.py,sha256=tvZ8ebod0VB1gCI6Pf4kkiUWrmA31DGVE_9E2MAjYXo,1743
 tico/pt2_to_circle.py,sha256=PPmFNw20jw2Z2VyM3ln9pX__jTzBOAZiv0gT5a-p-Y8,2666
 tico/config/__init__.py,sha256=xZzCXjZ84qE-CsBi-dfaL05bqpQ3stKKfTXhnrJRyVs,142
 tico/config/base.py,sha256=anwOiJFkUxUi7Cef573JgQcjk6S-FSi6O_TLjYASW-g,1244
@@ -54,7 +54,8 @@ tico/experimental/quantization/passes/fold_quant_ops.py,sha256=OYUndHaSL3GiL8koI
 tico/experimental/quantization/passes/insert_quantize_on_dtype_mismatch.py,sha256=AbNcI7rfIwHsQna_rFuwqFdOzFAU2lIB3sMK-vns8Dc,13072
 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/remove_weight_dequant_op.py,sha256=ZIwzuRXyCg6xN5pEUVV1LXsZpu0NkTcTUI6gc1VTuXs,6452
+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/interpreter.py,sha256=tGbluCbrehTCqBu8mtGDNzby_ieJ2ry8_RH_eC0CQxk,3828
@@ -134,6 +135,7 @@ tico/serialize/operators/op_gt.py,sha256=JAVbtuAUNLYhtJycJJCEkYo9QAvmiK4lTMdw5yH
 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_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
@@ -141,7 +143,7 @@ tico/serialize/operators/op_logical_and.py,sha256=WhQ8knuq32BO-WhAqkOgpcUStPkjoP
 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=SPXIdlrYgJfR982il_ajokNehFYJWOmccUoksS_0qos,5665
+tico/serialize/operators/op_max_pool2d_with_indices.py,sha256=Gd16ebLyo1x_gSXCvq-cBJEQ9jLTRLTd2zBXO7p3DvY,5724
 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
@@ -178,7 +180,7 @@ tico/serialize/operators/op_view.py,sha256=5EMww-ve17Vm9XPuV03Tn7vJsjpU2J8U4d_FO
 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=nbpBJue9_ezb8Dj9L9dWcNeVlZd0soGJu_fvrL6-c5c,11810
+tico/utils/convert.py,sha256=gh_EcWwjZZ5IyQKtazfiZ1_FvAMXe0leu5nPN-0OcHg,11919
 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
@@ -191,14 +193,14 @@ tico/utils/register_custom_op.py,sha256=iRQvdqlBqrJxq_pNkvJyDIJD_SYtCUl88wwbbuvS
 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=ZX820o05hmzg3QmmduAy-rIOrDyPHYe6Uj6Yf6Hoe_c,25055
+tico/utils/validate_args_kwargs.py,sha256=vp1GaGhOWqQBD4axMPaULy0tWxoIDTwEqEYoR5S7ph0,25244
 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.dev250617.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
-tico-0.1.0.dev250617.dist-info/METADATA,sha256=u9zQXsDpGwxI1Q2-6raMnRuzqsI0PIDGLjmq0mV3HjM,8846
-tico-0.1.0.dev250617.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
-tico-0.1.0.dev250617.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
-tico-0.1.0.dev250617.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
-tico-0.1.0.dev250617.dist-info/RECORD,,
+tico-0.1.0.dev250619.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
+tico-0.1.0.dev250619.dist-info/METADATA,sha256=Ms7oDZDD9txOcAycr9TUsJIGoC_B5FmXyF-q2AfJu58,8846
+tico-0.1.0.dev250619.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
+tico-0.1.0.dev250619.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
+tico-0.1.0.dev250619.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
+tico-0.1.0.dev250619.dist-info/RECORD,,