tico 0.1.0.dev250813__py3-none-any.whl → 0.1.0.dev250814__py3-none-any.whl

tico/__init__.py

@@ -29,7 +29,7 @@ __all__ = [
 ]
 
 # THIS LINE IS AUTOMATICALLY GENERATED BY setup.py
-__version__ = "0.1.0.dev250813"
+__version__ = "0.1.0.dev250814"
 
 MINIMUM_SUPPORTED_VERSION = "2.5.0"
 SECURE_TORCH_VERSION = "2.6.0"

tico/experimental/quantization/ptq/observers/__init__.py

@@ -1,9 +1,13 @@
 from tico.experimental.quantization.ptq.observers.affine_base import AffineObserverBase
 from tico.experimental.quantization.ptq.observers.base import ObserverBase
+from tico.experimental.quantization.ptq.observers.ema import EMAObserver
+from tico.experimental.quantization.ptq.observers.identity import IdentityObserver
 from tico.experimental.quantization.ptq.observers.minmax import MinMaxObserver
 
 __all__ = [
     "AffineObserverBase",
     "ObserverBase",
+    "EMAObserver",
+    "IdentityObserver",
     "MinMaxObserver",
 ]

tico/experimental/quantization/ptq/observers/ema.py

@@ -0,0 +1,62 @@
+# 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.
+
+import torch
+
+from tico.experimental.quantization.ptq.observers.affine_base import AffineObserverBase
+from tico.experimental.quantization.ptq.utils import channelwise_minmax
+
+
+class EMAObserver(AffineObserverBase):
+    """
+    Exponential-Moving-Average min/max tracker.
+
+    Why?
+    -----
+    • Smoother than raw MinMax (reduces outlier shock).
+    • Much cheaper than histogram/MSE observers.
+
+    The update rule follows the common "momentum" form:
+
+        ema = momentum * ema + (1 - momentum) * new_value
+
+    With momentum → 0: *fast* adaptation, momentum → 1: *slow* adaptation.
+    """
+
+    def __init__(
+        self,
+        *,
+        momentum: float = 0.9,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        assert 0.0 < momentum < 1.0, "momentum must be in (0, 1)"
+        self.momentum = momentum
+
+    @torch.no_grad()
+    def _update_stats(self, x: torch.Tensor):
+        if self.channel_axis is None:
+            curr_min, curr_max = x.min(), x.max()
+        else:
+            curr_min, curr_max = channelwise_minmax(x, self.channel_axis)
+
+        if (
+            torch.isinf(self.min_val).any() and torch.isinf(self.max_val).any()
+        ):  # first batch → hard init
+            self.min_val, self.max_val = curr_min, curr_max
+            return
+
+        m = self.momentum
+        self.min_val = m * self.min_val + (1 - m) * curr_min
+        self.max_val = m * self.max_val + (1 - m) * curr_max

tico/experimental/quantization/ptq/observers/identity.py

@@ -0,0 +1,74 @@
+# 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.
+
+"""
+IdentityObserver: a *no-op* observer for FP-only modules.
+
+Motivation
+----------
+Some layers should stay in full precision even when the rest of the model
+is quantized.  Attaching an `IdentityObserver` satisfies the wrapper API
+(`_update_stats()`, `compute_qparams()`, `fake_quant()`) without actually
+performing any statistics gathering or fake-quantization.
+"""
+import torch
+
+from tico.experimental.quantization.ptq.observers.affine_base import AffineObserverBase
+
+
+class IdentityObserver(AffineObserverBase):
+    """
+    Passthrough observer that **never** alters the tensor.
+
+    • `_update_stats()`   → does nothing
+    • `compute_qparams()` → returns (1.0, 0) *dummy* q-params
+    • `fake_quant()`      → returns *x* unchanged
+    """
+
+    def __init__(self, **kwargs):
+        # Call parent so the usual fields (`dtype`, `qscheme`, …) exist,
+        # but immediately disable any stateful behaviour.
+        super().__init__(**kwargs)
+
+        # Deactivate statistics collection permanently.
+        self.enabled = False
+
+        # Pre-cache sentinel q-params so wrapper code that blindly
+        # accesses them won't crash.
+        self._cached_scale = torch.tensor(1.0)
+        self._cached_zp = torch.tensor(0, dtype=torch.int)
+
+    def reset(self) -> None:  # (simple override – nothing to do)
+        """No internal state to reset."""
+        pass
+
+    def _update_stats(self, x: torch.Tensor) -> None:
+        """Skip statistic collection entirely."""
+        return
+
+    def compute_qparams(self):
+        """
+        Return the pre-cached (scale, zero_point) tuple.
+
+        Keeping the signature identical to other observers allows uniform
+        lifecycle management in wrapper code.
+        """
+        return self._cached_scale, self._cached_zp
+
+    def fake_quant(self, x: torch.Tensor):
+        """Identity mapping — leaves *x* in FP."""
+        return x
+
+    def __repr__(self) -> str:
+        return f"{self.__class__.__name__}()"

tico/utils/utils.py

@@ -79,73 +79,70 @@ def enforce_type(callable):
     def check_types(*args, **kwargs):
         parameters = dict(zip(spec.args, args))
         parameters.update(kwargs)
-        for name, value in parameters.items():
-            if name == "self":
-                # skip 'self' in spec.args
-                continue
-
-            assert (
-                name in spec.annotations
-            ), f"All parameter require type hints. {name} needs a type hint"
-
-            type_hint = spec.annotations[name]
 
-            # Return tuple of flattened types.
-            # Q) What is flatten?
-            # A) Optional/Union is not included. Below are included.
-            # collections: List, Set, ...
-            # primitive types: int, str, ...
-            def _flatten_type(type_hint) -> tuple:
-                # `get_origin` maps Union[...] and Optional[...] varieties to Union
-                if typing.get_origin(type_hint) == typing.Union:
-                    # ex. typing.Union[list, int] -> (list, int)
-                    # ex. typing.Optional[torch.fx.Node] -> (torch.fx.Node, NoneType)
-                    actual_type = tuple(
-                        [_flatten_type(t) for t in typing.get_args(type_hint)]
-                    )
-                else:
-                    actual_type = (type_hint,)
-                return actual_type
+        # Return tuple of flattened types.
+        # Q) What is flatten?
+        # A) Optional/Union is not included. Below are included.
+        # collections: List, Set, ...
+        # primitive types: int, str, ...
+        def _flatten_type(type_hint) -> tuple:
+            # `get_origin` maps Union[...] and Optional[...] varieties to Union
+            if typing.get_origin(type_hint) == typing.Union:
+                # ex. typing.Union[list, int] -> (list, int)
+                # ex. typing.Optional[torch.fx.Node] -> (torch.fx.Node, NoneType)
+                actual_type = tuple(
+                    _flatten_type(t) for t in typing.get_args(type_hint)
+                )
+            else:
+                actual_type = (type_hint,)
+            return actual_type
 
-            type_hint = _flatten_type(type_hint)
+        # Return true if value matches with type_hint
+        # Return false otherwise
+        def _check_type(value, type_hint):
+            if type_hint == typing.Any:
+                return True
 
-            # Return true if value matches with type_hint
-            # Return false otherwise
-            def _check_type(value, type_hint):
-                if type_hint == typing.Any:
-                    return True
+            if isinstance(type_hint, tuple):
+                return any(_check_type(value, t) for t in type_hint)
 
-                if isinstance(type_hint, tuple):
-                    return any([_check_type(value, t) for t in type_hint])
+            if typing.get_origin(type_hint) in (list, set):
+                if not isinstance(value, typing.get_origin(type_hint)):
+                    return False
 
-                if typing.get_origin(type_hint) in (list, set):
-                    if not isinstance(value, typing.get_origin(type_hint)):
+                for v in value:
+                    if not any(_check_type(v, t) for t in typing.get_args(type_hint)):
                         return False
 
-                    for v in value:
-                        if not any(
-                            [_check_type(v, t) for t in typing.get_args(type_hint)]
-                        ):
-                            return False
+                return True
 
-                    return True
+            if typing.get_origin(type_hint) is dict:
+                if not isinstance(value, typing.get_origin(type_hint)):
+                    return False
 
-                if typing.get_origin(type_hint) is dict:
-                    if not isinstance(value, typing.get_origin(type_hint)):
+                for k, v in value.items():
+                    k_type, v_type = typing.get_args(type_hint)
+                    if not _check_type(k, k_type):
+                        return False
+                    if not _check_type(v, v_type):
                         return False
 
-                    for k, v in value.items():
-                        k_type, v_type = typing.get_args(type_hint)
-                        if not _check_type(k, k_type):
-                            return False
-                        if not _check_type(v, v_type):
-                            return False
+                return True
+
+            # TODO: Support more type hints
+            return isinstance(value, type_hint)
 
-                    return True
+        for name, value in parameters.items():
+            if name == "self":
+                # skip 'self' in spec.args
+                continue
 
-                # TODO: Support more type hints
-                return isinstance(value, type_hint)
+            assert (
+                name in spec.annotations
+            ), f"All parameter require type hints. {name} needs a type hint"
 
+            type_hint = spec.annotations[name]
+            type_hint = _flatten_type(type_hint)
             type_check_result = _check_type(value, type_hint)
             if not type_check_result:
                 raise ArgTypeError(

{tico-0.1.0.dev250813.dist-info → tico-0.1.0.dev250814.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: tico
-Version: 0.1.0.dev250813
+Version: 0.1.0.dev250814
 Summary: Convert exported Torch module to circle
 Home-page: UNKNOWN
 License: UNKNOWN

{tico-0.1.0.dev250813.dist-info → tico-0.1.0.dev250814.dist-info}/RECORD

@@ -1,4 +1,4 @@
-tico/__init__.py,sha256=SL1CaKezytxIZbCthWmNK0MGO-5Erit1aRwOxaT6SxU,1883
+tico/__init__.py,sha256=Y5AdGv7QfIPC6o_P2M2S16JtXG5LLmtIfUU7gq6scYQ,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
@@ -60,9 +60,11 @@ tico/experimental/quantization/ptq/__init__.py,sha256=ZoPdEwZ1i1n5pBFChx8GuUrkfR
 tico/experimental/quantization/ptq/dtypes.py,sha256=xfCBtq6mQmUYRwsoFgII6gvRl1raQi0Inj9pznDuKwQ,2236
 tico/experimental/quantization/ptq/mode.py,sha256=lT-T8vIv8YWcwrjT7xXVhOw1g7aoAdh_3PWB-ptPKaI,1052
 tico/experimental/quantization/ptq/qscheme.py,sha256=uwhv7bCxOOXB3I-IKlRyr_u4eXOq48uIqGy4TLDqGxY,1301
-tico/experimental/quantization/ptq/observers/__init__.py,sha256=Sxdr1xAKhmdUfwwzsNfer0JVGiT3Xo6I1Bw3TRkp3C0,325
+tico/experimental/quantization/ptq/observers/__init__.py,sha256=wyrO0KTZve78aFWTwvsOE82Vu2kbCxJv8aqjiO1QL2s,524
 tico/experimental/quantization/ptq/observers/affine_base.py,sha256=e2Eba64nrxKQyE4F_WJ7WTSsk3xe6bkdGUKaoLFWGFw,4638
 tico/experimental/quantization/ptq/observers/base.py,sha256=Wons1MzpqK1mfcy-ppl-B2Dum0edXg2dWW2Lw3V18tw,3280
+tico/experimental/quantization/ptq/observers/ema.py,sha256=WZiYWEHrkgizAwnRCtfOm9JPHfZrjZTxMr6X9Wuovmo,2061
+tico/experimental/quantization/ptq/observers/identity.py,sha256=jdlNH52z8ANOZbs_0KFZ4iEstVfNC1OUzQsm1a9FFpM,2595
 tico/experimental/quantization/ptq/observers/minmax.py,sha256=mLHkwIzWFzQXev7EU7w1333KckwRjukc3_cUPJOnUfs,1486
 tico/experimental/quantization/ptq/utils/__init__.py,sha256=PL9IZgiWoMtsXVljeOy7KymmLVP238SXEFRLXYK72WQ,126
 tico/experimental/quantization/ptq/utils/reduce_utils.py,sha256=3kWawLB91EcvvHlCrNqqfZF7tpgr22htBSA049mKw_4,973
@@ -214,15 +216,15 @@ tico/utils/serialize.py,sha256=mEuusEzi82WFsz3AkowgWwxSLeo50JDxyOj6yYDQhEI,1914
 tico/utils/signature.py,sha256=R2GV0alRpXEbZISqPKyxCUWbgDcsrQ2ovbVG3737IzA,9595
 tico/utils/torch_compat.py,sha256=oc6PztVsXdHcQ3iaVR90wLLxrGaj6zFHWZ8K9rRS6q8,1795
 tico/utils/trace_decorators.py,sha256=ddLIiKQfSaQrxgF1kNpwjFTQnXENzeSfcr1kuAW4jGI,3221
-tico/utils/utils.py,sha256=A5p3iAAxRGDsZJh4ybp-Qo3MX3vk5RrmSY-R3rXqVeI,12976
+tico/utils/utils.py,sha256=aySftYnNTsqVAMcGs_3uX3-hz577a2cj4p1aVV-1XeQ,12747
 tico/utils/validate_args_kwargs.py,sha256=aY7hyDaZKrZn0ev0liUFZdHS8__0Vpp5QyqDEobZ_zM,27163
 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.dev250813.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
-tico-0.1.0.dev250813.dist-info/METADATA,sha256=CYADXCywzXQsqDipD-SWVOIKjxDRTaWHL1tv_LuQAkA,8450
-tico-0.1.0.dev250813.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
-tico-0.1.0.dev250813.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
-tico-0.1.0.dev250813.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
-tico-0.1.0.dev250813.dist-info/RECORD,,
+tico-0.1.0.dev250814.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
+tico-0.1.0.dev250814.dist-info/METADATA,sha256=CWbuRvOwPOiqf8FwnG5H8tij1fQqH3k_JpO8zhtXQfg,8450
+tico-0.1.0.dev250814.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
+tico-0.1.0.dev250814.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
+tico-0.1.0.dev250814.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
+tico-0.1.0.dev250814.dist-info/RECORD,,