tico 0.1.0.dev250812__py3-none-any.whl → 0.1.0.dev250813__py3-none-any.whl

tico/__init__.py

@@ -29,7 +29,7 @@ __all__ = [
 ]
 
 # THIS LINE IS AUTOMATICALLY GENERATED BY setup.py
-__version__ = "0.1.0.dev250812"
+__version__ = "0.1.0.dev250813"
 
 MINIMUM_SUPPORTED_VERSION = "2.5.0"
 SECURE_TORCH_VERSION = "2.6.0"

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

@@ -0,0 +1,9 @@
+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.minmax import MinMaxObserver
+
+__all__ = [
+    "AffineObserverBase",
+    "ObserverBase",
+    "MinMaxObserver",
+]

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

@@ -0,0 +1,128 @@
+# 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 math
+from typing import Optional, Tuple
+
+import torch
+
+from tico.experimental.quantization.ptq.dtypes import DType, UINT8
+from tico.experimental.quantization.ptq.observers.base import ObserverBase
+from tico.experimental.quantization.ptq.qscheme import QScheme
+
+
+class AffineObserverBase(ObserverBase):
+    """Base for affine observers (min/max → scale/zp)."""
+
+    def __init__(
+        self,
+        *,
+        name: str,
+        dtype: DType = UINT8,
+        qscheme: QScheme = QScheme.PER_TENSOR_ASYMM,
+        channel_axis: Optional[int] = None,
+    ):
+        super().__init__(
+            name=name, dtype=dtype, qscheme=qscheme, channel_axis=channel_axis
+        )
+
+    def reset(self) -> None:
+        """
+        Reset running min/max and drop cached qparams.
+        """
+        self.min_val: torch.Tensor = torch.tensor(math.inf)
+        self.max_val: torch.Tensor = torch.tensor(-math.inf)
+        if hasattr(self, "_cached_scale"):
+            del self._cached_scale
+        if hasattr(self, "_cached_zp"):
+            del self._cached_zp
+
+    def load_qparams(self, scale: torch.Tensor, zp: torch.Tensor, *, lock: bool = True):
+        """
+        Inject externally computed qparams and optionally lock the observer.
+
+        When locked, subsequent `collect()` calls are ignored.
+        """
+        self._cached_scale = scale.detach()
+        self._cached_zp = zp.to(torch.int)
+        if lock:
+            self.enabled = False
+
+    @property
+    def has_qparams(self) -> bool:
+        return hasattr(self, "_cached_scale")
+
+    def compute_qparams(self) -> Tuple[torch.Tensor, torch.Tensor]:
+        qmin, qmax = self.dtype.qmin, self.dtype.qmax
+        rng = self.max_val - self.min_val
+        eps = 1e-12
+
+        if self.qscheme.is_symmetric():
+            max_abs = torch.maximum(self.max_val.abs(), self.min_val.abs())
+            scale = torch.clamp(max_abs, min=eps) / qmax
+            zp = torch.zeros_like(scale, dtype=torch.int)
+            self._cached_scale, self._cached_zp = scale, zp
+            return scale, zp
+
+        if self.channel_axis is None:
+            if torch.all(rng.abs() < 1e-8):
+                C = self.min_val
+                if torch.allclose(C, torch.zeros_like(C)):
+                    scale = torch.ones_like(C)
+                    zp = torch.zeros_like(C, dtype=torch.int)
+                elif (C > 0).all():
+                    scale = torch.clamp(C, min=eps)
+                    zp = torch.zeros_like(C, dtype=torch.int)
+                else:
+                    scale = torch.clamp(C.abs(), min=eps)
+                    zp = torch.full_like(C, qmax, dtype=torch.int)
+            else:
+                scale = torch.clamp(rng, min=eps) / (qmax - qmin)
+                zp = (
+                    torch.round(qmin - self.min_val / scale)
+                    .clamp(qmin, qmax)
+                    .to(torch.int)
+                )
+        else:
+            scale = torch.clamp(rng, min=eps) / (qmax - qmin)
+            zp = (
+                torch.round(qmin - self.min_val / scale).clamp(qmin, qmax).to(torch.int)
+            )
+
+        self._cached_scale, self._cached_zp = scale, zp
+        return scale, zp
+
+    def fake_quant(self, x: torch.Tensor) -> torch.Tensor:
+        if not self.has_qparams:
+            raise RuntimeError(
+                "Call compute_qparams()/freeze_qparams() or load_qparams() first."
+            )
+        scale, zp = self._cached_scale, self._cached_zp
+        if self.channel_axis is None:
+            return torch.fake_quantize_per_tensor_affine(
+                x,
+                scale=scale,
+                zero_point=zp,
+                quant_min=self.dtype.qmin,
+                quant_max=self.dtype.qmax,
+            )
+        else:
+            return torch.fake_quantize_per_channel_affine(
+                x,
+                scale=scale,
+                zero_point=zp,
+                axis=self.channel_axis,
+                quant_min=self.dtype.qmin,
+                quant_max=self.dtype.qmax,
+            )

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

@@ -0,0 +1,98 @@
+# 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 abc import ABC, abstractmethod
+from typing import Optional, Tuple
+
+import torch
+
+from tico.experimental.quantization.ptq.dtypes import DType, UINT8
+from tico.experimental.quantization.ptq.qscheme import QScheme
+
+
+class ObserverBase(ABC):
+    """
+    Minimal abstract base for all observers/quantizers.
+
+    Subclasses must implement:
+      - reset()
+      - collect(x)
+      - fake_quant(x)
+      - compute_qparams(): optional in practice for some observers (e.g., MX),
+        but still part of the interface; those can return None.
+    """
+
+    def __init__(
+        self,
+        *,
+        name: str,
+        dtype: DType = UINT8,
+        qscheme: QScheme = QScheme.PER_TENSOR_ASYMM,
+        channel_axis: Optional[int] = None,  # None → per-tensor
+    ):
+        self.name = name
+        self.dtype = dtype
+        self.qscheme = qscheme
+        self.channel_axis = channel_axis if qscheme.is_per_channel() else None
+        self.enabled = True
+        self.reset()
+
+    @abstractmethod
+    def reset(self) -> None:
+        """Clear any running statistics or cached params."""
+        raise NotImplementedError
+
+    def collect(self, x: torch.Tensor) -> None:
+        """
+        Update running statistics with a new batch of data.
+
+        This base implementation guards on `enabled` and then calls `_update_stats(x)`.
+        Subclasses should implement `_update_stats(x)` instead of overriding `collect`.
+        """
+        if not self.enabled:
+            return
+        self._update_stats(x)
+
+    @abstractmethod
+    def _update_stats(self, x: torch.Tensor) -> None:
+        """
+        Update running statistics (min/max, hist, mse buffers, ...).
+
+        Must be implemented by subclasses (e.g., MinMax, EMA, Histogram, MSE).
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def fake_quant(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Apply the observer's quantization.
+        Implementations may or may not rely on qparams.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def compute_qparams(self) -> Optional[Tuple[torch.Tensor, torch.Tensor]]:
+        """
+        Compute and (if applicable) cache quantization params.
+        Affine observers typically return (scale, zero_point).
+        Observers that do not use qparams (e.g., MX) may return None.
+        """
+        raise NotImplementedError
+
+    # String repr helps debugging
+    def __repr__(self) -> str:
+        return (
+            f"{self.__class__.__name__}(name={self.name}, dtype={str(self.dtype)}, "
+            f"qscheme={str(self.qscheme)}, channel_axis={self.channel_axis}, enabled={self.enabled})"
+        )

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

@@ -0,0 +1,39 @@
+# 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 MinMaxObserver(AffineObserverBase):
+    """Plain min/max range tracker."""
+
+    @torch.no_grad()
+    def _update_stats(self, x: torch.Tensor) -> None:
+        """
+        Update running min/max with the incoming batch.
+
+        Per-tensor: use global min/max.
+        Per-channel: reduce all axes except the channel axis.
+        """
+        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)
+
+        # Broadcasting handles scalar-vs-vector cases
+        self.min_val = torch.minimum(self.min_val, curr_min)
+        self.max_val = torch.maximum(self.max_val, curr_max)

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

@@ -0,0 +1,5 @@
+from tico.experimental.quantization.ptq.utils.reduce_utils import channelwise_minmax
+
+__all__ = [
+    "channelwise_minmax",
+]

tico/experimental/quantization/ptq/utils/reduce_utils.py

@@ -0,0 +1,25 @@
+# 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
+
+
+def channelwise_minmax(x: torch.Tensor, channel_axis: int):
+    """
+    Compute per-channel (min, max) by reducing all axes except `channel_axis`.
+    """
+    channel_axis = channel_axis % x.ndim  # handle negative indices safely
+    dims = tuple(d for d in range(x.ndim) if d != channel_axis)
+
+    return x.amin(dim=dims), x.amax(dim=dims)

tico/passes/decompose_fake_quantize_tensor_qparams.py

@@ -244,7 +244,7 @@ class DecomposeFakeQuantizeTensorQParams(PassBase):
                     # So, let's remove `mask` from the output.args first.
                     # mask_user(output).args == (dequantize_per_tensor.tensor, mask)
                     if mask:
-                        len(mask) == 1
+                        assert len(mask) == 1
                         mask_user = list(mask[0].users.keys())[0]
                         assert len(mask_user.args) == 1
                         mask_user.args = ((mask_user.args[0][0],),)

tico/utils/register_custom_op.py

@@ -707,7 +707,7 @@ def CircleRMSNorm():
     @custom_op("circle_custom::rms_norm", mutates_args=())
     def rms_norm(
         hidden_states: torch.Tensor,
-        weight: Optional[torch.Tensor] = None,
+        weight: torch.Tensor,
         eps: float = 1e-05,
     ) -> torch.Tensor:
         input_dtype = hidden_states.dtype
@@ -719,7 +719,7 @@ def CircleRMSNorm():
     @register_fake("circle_custom::rms_norm")
     def _(
         hidden_states: torch.Tensor,
-        weight: Optional[torch.Tensor] = None,
+        weight: torch.Tensor,
         eps: float = 1e-05,
     ) -> torch.Tensor:
         return hidden_states.new_empty(hidden_states.size())

tico/utils/validate_args_kwargs.py

@@ -175,13 +175,12 @@ class CatArgs:
 @dataclass
 class CircleRMSNormArgs:
     """
-    This is not aten ops but custom op for RMSNorm.
-    circle_custom.rms_norm(Tensor input, Tensor? weight=None, float? eps=None) -> Tensor
+    For circle.BuiltinOperator.BuiltinOperator.RMS_NORM
     """
 
     input: torch.fx.Node
-    weight: Optional[torch.fx.Node]
-    eps: Optional[float]
+    weight: torch.fx.Node
+    eps: float
 
 
 @enforce_type

{tico-0.1.0.dev250812.dist-info → tico-0.1.0.dev250813.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: tico
-Version: 0.1.0.dev250812
+Version: 0.1.0.dev250813
 Summary: Convert exported Torch module to circle
 Home-page: UNKNOWN
 License: UNKNOWN

{tico-0.1.0.dev250812.dist-info → tico-0.1.0.dev250813.dist-info}/RECORD

@@ -1,4 +1,4 @@
-tico/__init__.py,sha256=QzZEOB1tLI8yPD6dzOftGhdVFbFkAintnJP1vv89AJQ,1883
+tico/__init__.py,sha256=SL1CaKezytxIZbCthWmNK0MGO-5Erit1aRwOxaT6SxU,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
@@ -60,6 +60,12 @@ 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/affine_base.py,sha256=e2Eba64nrxKQyE4F_WJ7WTSsk3xe6bkdGUKaoLFWGFw,4638
+tico/experimental/quantization/ptq/observers/base.py,sha256=Wons1MzpqK1mfcy-ppl-B2Dum0edXg2dWW2Lw3V18tw,3280
+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
 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
@@ -75,7 +81,7 @@ tico/passes/convert_to_relu6.py,sha256=1BJpUwUb6Zli_1y3eyJQo7dg9B1xvZ7sYjMbvEQsF
 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
-tico/passes/decompose_fake_quantize_tensor_qparams.py,sha256=k9MJhMVABFNF6lXgEum1fJyGpdQwVRKxWOYhkMR2M7c,13915
+tico/passes/decompose_fake_quantize_tensor_qparams.py,sha256=p-sz_cgir4jMWp43CR75fj0TbLkNvSl888fvkRqFRtE,13922
 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
@@ -203,20 +209,20 @@ tico/utils/padding.py,sha256=qKke-dJeeLHiRaePjDS66txrGyiYuipLVQeqLYad8uk,3349
 tico/utils/passes.py,sha256=kGmDe__5cPaO6i5EDAoXSVe6yXEoX9hAny4ROb3ZEmQ,2409
 tico/utils/pytree_utils.py,sha256=jrk3N6X6LiUnBCX_gM1K9nywbVAJBVnszlTAgeIeDUc,5219
 tico/utils/record_input.py,sha256=QN-8D71G_WAX3QQQ5CIwbEfFJZTQ3CvL4wCMiVddua4,3894
-tico/utils/register_custom_op.py,sha256=n91UtmPedoqhkR8fBNRbk9Msq79pn9DHNHlt99l2s_w,28142
+tico/utils/register_custom_op.py,sha256=dPemLyjrf4xMYCUlhhZeUhixL8Eat3Ywlv6K5kTqG8Y,28108
 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/validate_args_kwargs.py,sha256=yikeUbYfSg2378wagEMXDlJeSRv8HKI2oxpjWarolec,27268
+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.dev250812.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
-tico-0.1.0.dev250812.dist-info/METADATA,sha256=VYlCGN-A_iI52ctvoTcdzg3UrtQ2zhZRCxyunaCdiqU,8450
-tico-0.1.0.dev250812.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
-tico-0.1.0.dev250812.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
-tico-0.1.0.dev250812.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
-tico-0.1.0.dev250812.dist-info/RECORD,,
+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,,