tico 0.1.0.dev250819__py3-none-any.whl → 0.1.0.dev250821__py3-none-any.whl

tico/__init__.py

@@ -29,7 +29,7 @@ __all__ = [
 ]
 
 # THIS LINE IS AUTOMATICALLY GENERATED BY setup.py
-__version__ = "0.1.0.dev250819"
+__version__ = "0.1.0.dev250821"
 
 MINIMUM_SUPPORTED_VERSION = "2.5.0"
 SECURE_TORCH_VERSION = "2.6.0"

tico/experimental/quantization/ptq/wrappers/nn/__init__.py

@@ -0,0 +1,5 @@
+from tico.experimental.quantization.ptq.wrappers.nn.quant_linear import QuantLinear
+
+__all__ = [
+    "QuantLinear",
+]

tico/experimental/quantization/ptq/wrappers/nn/quant_linear.py

@@ -0,0 +1,66 @@
+# 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 Optional
+
+import torch.nn as nn
+import torch.nn.functional as F
+
+from tico.experimental.quantization.ptq.mode import Mode
+from tico.experimental.quantization.ptq.qscheme import QScheme
+from tico.experimental.quantization.ptq.quant_config import QuantConfig
+from tico.experimental.quantization.ptq.wrappers.quant_module_base import (
+    QuantModuleBase,
+)
+from tico.experimental.quantization.ptq.wrappers.registry import register
+
+
+@register(nn.Linear)
+class QuantLinear(QuantModuleBase):
+    """Per-channel weight fake-quant, eager-output activation fake-quant."""
+
+    def __init__(
+        self,
+        fp: nn.Linear,
+        *,
+        qcfg: Optional[QuantConfig] = None,
+        fp_name: Optional[str] = None
+    ):
+        super().__init__(qcfg, fp_name=fp_name)
+        self.weight_obs = self._make_obs(
+            "weight", qscheme=QScheme.PER_CHANNEL_ASYMM, channel_axis=0
+        )
+        self.act_in_obs = self._make_obs("act_in")
+        self.act_out_obs = self._make_obs("act_out")
+        self.module = fp
+
+    def enable_calibration(self) -> None:
+        super().enable_calibration()
+        # immediately capture the fixed weight range
+        self.weight_obs.collect(self.module.weight)
+
+    def forward(self, x):
+        x_q = self._fq(x, self.act_in_obs)
+
+        w = self.module.weight
+        if self._mode is Mode.QUANT:
+            w = self.weight_obs.fake_quant(w)
+        b = self.module.bias
+
+        out = F.linear(x_q, w, b)
+
+        return self._fq(out, self.act_out_obs)
+
+    def _all_observers(self):
+        return (self.weight_obs, self.act_in_obs, self.act_out_obs)

tico/experimental/quantization/ptq/wrappers/ptq_wrapper.py

@@ -0,0 +1,54 @@
+# 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 Optional
+
+import torch
+
+from tico.experimental.quantization.ptq.quant_config import QuantConfig
+from tico.experimental.quantization.ptq.wrappers.quant_module_base import (
+    QuantModuleBase,
+)
+from tico.experimental.quantization.ptq.wrappers.registry import lookup
+
+
+class PTQWrapper(QuantModuleBase):
+    """
+    Adapter that turns a fp module into its quantized counterpart.
+
+    It is itself a QuantModuleBase so composite wrappers can treat
+     it exactly like any other quant module.
+    """
+
+    def __init__(
+        self,
+        module: torch.nn.Module,
+        qcfg: Optional[QuantConfig] = None,
+        *,
+        fp_name: Optional[str] = None,
+    ):
+        super().__init__(qcfg)
+        wrapped_cls = lookup(type(module))
+        if wrapped_cls is None:
+            raise NotImplementedError(f"No quant wrapper for {type(module).__name__}")
+        self.wrapped: QuantModuleBase = wrapped_cls(module, qcfg=qcfg, fp_name=fp_name)  # type: ignore[arg-type, misc]
+
+    def forward(self, *args, **kwargs):
+        return self.wrapped(*args, **kwargs)
+
+    def _all_observers(self):
+        yield from self.wrapped._all_observers()
+
+    def extra_repr(self) -> str:
+        return self.wrapped.extra_repr()

tico/experimental/quantization/ptq/wrappers/quant_elementwise.py

@@ -0,0 +1,113 @@
+# 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 Callable, Optional
+
+import torch
+import torch.nn as nn
+
+from tico.experimental.quantization.ptq.quant_config import QuantConfig
+from tico.experimental.quantization.ptq.wrappers.quant_module_base import (
+    QuantModuleBase,
+)
+from tico.experimental.quantization.ptq.wrappers.registry import register
+
+
+class QuantElementwise(QuantModuleBase):
+    """
+    Generic wrapper for any 1-to-1 element-wise op `y = f(x)`.
+
+    Sub-classes only need to implement:
+        • `FUNC`: a Callable that maps tensor→tensor
+    """
+
+    # subclass must set this
+    FUNC: Callable[[torch.Tensor], torch.Tensor] | None = None
+
+    def __init_subclass__(cls, **kwargs):
+        super().__init_subclass__(**kwargs)
+        if cls is QuantElementwise:
+            return
+        if cls.FUNC is None:
+            raise NotImplementedError(
+                f"{cls.__name__} must define a staticmethod `FUNC(tensor) -> tensor`"
+            )
+
+    def __init__(
+        self,
+        fp_module: nn.Module,
+        *,
+        qcfg: Optional[QuantConfig] = None,
+        fp_name: Optional[str] = None,
+    ):
+        super().__init__(qcfg, fp_name=fp_name)
+        self.module = fp_module
+        self.act_in_obs = self._make_obs("act_in")
+        self.act_out_obs = self._make_obs("act_out")
+
+    # ------------------------------------------------------------
+    def forward(self, x):
+        x_q = self._fq(x, self.act_in_obs)
+        assert self.FUNC is not None
+        y = self.FUNC(x_q)  # element-wise op
+        y_q = self._fq(y, self.act_out_obs)
+        return y_q
+
+    # ------------------------------------------------------------
+    def _all_observers(self):
+        return (self.act_in_obs, self.act_out_obs)
+
+
+"""
+Why `FUNC` is a `staticmethod`
+
+- Prevents automatic binding: calling `self.FUNC(x)` will not inject `self`,
+  so the callable keeps the expected signature `Tensor -> Tensor`
+  (e.g., `torch.sigmoid(x)`), avoiding TypeErrors.
+
+- Expresses purity and statelessness: `FUNC` is a pure, element-wise transform
+  that must not read or mutate module state (params, buffers, config).
+
+- Tracing/export friendly (FX / TorchScript): the call is captured as
+  `call_function(torch.*)` instead of a bound `call_method`, which makes graph
+  rewrites/pattern-matching and backends' substitutions more reliable.
+
+- Avoids submodule pollution: we keep a functional op (`torch.relu`) rather
+  than an `nn.Module` instance that would appear in the module tree.
+
+- Small perf/alloc win: no bound-method objects are created on each call.
+"""
+
+# Sigmoid
+@register(nn.Sigmoid)
+class QuantSigmoid(QuantElementwise):
+    FUNC = staticmethod(torch.sigmoid)
+
+
+# Tanh
+@register(nn.Tanh)
+class QuantTanh(QuantElementwise):
+    FUNC = staticmethod(torch.tanh)
+
+
+# ReLU
+@register(nn.ReLU)
+class QuantReLU(QuantElementwise):
+    FUNC = staticmethod(torch.relu)
+
+
+# GELU (approximate)
+@register(nn.GELU)
+class QuantGELU(QuantElementwise):
+    FUNC = staticmethod(torch.nn.functional.gelu)

tico/experimental/quantization/ptq/wrappers/registry.py

@@ -0,0 +1,112 @@
+# 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 importlib
+from typing import Callable, Dict, Type
+
+import torch.nn as nn
+
+from tico.experimental.quantization.ptq.wrappers.quant_module_base import (
+    QuantModuleBase,
+)
+
+_WRAPPERS: Dict[Type[nn.Module], Type[QuantModuleBase]] = {}
+_IMPORT_ONCE = False
+_CORE_MODULES = (
+    "tico.experimental.quantization.ptq.wrappers.nn.quant_linear",
+    # add future core wrappers here
+)
+
+
+def _lazy_init():
+    """
+    Deferred one-shot import of "core wrapper modules".
+
+    Why not import everything when the program first starts?
+    --------------------------------------------------
+    * **Avoid circular-import hell**
+      Core wrappers often import `PTQWrapper`, which in turn calls
+      `registry.lookup()`.  Importing those files eagerly here would create a
+      cycle (`registry → wrapper → registry`).  Delaying the import until the
+      *first* `lookup()` call lets Python finish constructing the registry
+      module before any wrapper files are touched.
+
+    * **Cold-start speed**
+      Most user code never wraps layers explicitly; they only hit
+      `PTQWrapper` if they are doing quantization.  Deferring half-a-dozen
+      heavyweight `import torch …` files until they are really needed
+      reduces library start-up latency in the common path.
+
+    * **Optional dependencies**
+      Core wrappers listed in `_CORE_MODULES` are chosen to be dependency-free
+      (pure PyTorch).  Anything that needs `transformers`, `torchvision`,
+      etc. uses the `@try_register()` decorator inside its own module.  Those
+      optional modules are *not* imported here, so users without the extra
+      packages still get a clean import.
+
+    Implementation notes
+    --------------------
+    * `_IMPORT_ONCE` guard ensures we execute the import loop only once,
+      even if `lookup()` is called from multiple threads.
+    * Each path in `_CORE_MODULES` is a "fully-qualified module string"
+      (e.g. "ptq.wrappers.linear_quant").  Importing the module runs all
+      its `@register(nn.Layer)` decorators, populating `_WRAPPERS`.
+    * After the first call the function becomes a cheap constant-time no-op.
+    """
+    global _IMPORT_ONCE
+    if _IMPORT_ONCE:
+        return
+    for mod in _CORE_MODULES:
+        __import__(mod)  # triggers decorators
+    _IMPORT_ONCE = True
+
+
+# ───────────────────────────── decorator for always-present classes
+def register(
+    fp_cls: Type[nn.Module],
+) -> Callable[[Type[QuantModuleBase]], Type[QuantModuleBase]]:
+    def _decorator(quant_cls: Type[QuantModuleBase]):
+        _WRAPPERS[fp_cls] = quant_cls
+        return quant_cls
+
+    return _decorator
+
+
+# ───────────────────────────── conditional decorator
+def try_register(path: str) -> Callable[[Type[QuantModuleBase]], Type[QuantModuleBase]]:
+    """
+    @try_register("transformers.models.llama.modeling_llama.LlamaMLP")
+
+    • If import succeeds → behave like `@register`
+    • If module/class not found → become a NO-OP
+    """
+
+    def _decorator(quant_cls: Type[QuantModuleBase]):
+        module_name, _, cls_name = path.rpartition(".")
+        try:
+            mod = importlib.import_module(module_name)
+            fp_cls = getattr(mod, cls_name)
+            _WRAPPERS[fp_cls] = quant_cls
+        except (ModuleNotFoundError, AttributeError):
+            # transformers not installed or class renamed – silently skip
+            pass
+        return quant_cls
+
+    return _decorator
+
+
+# ───────────────────────────── lookup
+def lookup(fp_cls: Type[nn.Module]) -> Type[QuantModuleBase] | None:
+    _lazy_init()
+    return _WRAPPERS.get(fp_cls)

{tico-0.1.0.dev250819.dist-info → tico-0.1.0.dev250821.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: tico
-Version: 0.1.0.dev250819
+Version: 0.1.0.dev250821
 Summary: Convert exported Torch module to circle
 Home-page: UNKNOWN
 License: UNKNOWN

{tico-0.1.0.dev250819.dist-info → tico-0.1.0.dev250821.dist-info}/RECORD

@@ -1,4 +1,4 @@
-tico/__init__.py,sha256=lsEhU5spgB7ZLxS5ZhEhaDXDrDhT-14eoev-lcvlIx0,1883
+tico/__init__.py,sha256=yEGf_7r1bSZi_LwntDyrVf0WUNvouR4X_fcua51LRUI,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
@@ -71,7 +71,12 @@ tico/experimental/quantization/ptq/observers/mx.py,sha256=aP4qmBgeiRIYZJksShN5gs
 tico/experimental/quantization/ptq/utils/__init__.py,sha256=PL9IZgiWoMtsXVljeOy7KymmLVP238SXEFRLXYK72WQ,126
 tico/experimental/quantization/ptq/utils/reduce_utils.py,sha256=3kWawLB91EcvvHlCrNqqfZF7tpgr22htBSA049mKw_4,973
 tico/experimental/quantization/ptq/wrappers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+tico/experimental/quantization/ptq/wrappers/ptq_wrapper.py,sha256=KRw_VvFJYvd2OBj4K1sYEXxUwZk9QghMw3NsgjKIAGk,1857
+tico/experimental/quantization/ptq/wrappers/quant_elementwise.py,sha256=LhEoobfvto6zKrBOKL4gmxfFFc31jHzyQV_zfps-iQM,3604
 tico/experimental/quantization/ptq/wrappers/quant_module_base.py,sha256=6RK4bn9G1pzFmkIdBdFf7liBOpb-b7rpthgD83AgkbQ,5256
+tico/experimental/quantization/ptq/wrappers/registry.py,sha256=exXl2wNNzVgC2P9gMjpF_-PqIBgYERGruzh0u1Pril0,4367
+tico/experimental/quantization/ptq/wrappers/nn/__init__.py,sha256=q4A9BiGlsa8ZdGV3y0SDiSkzkdVugsK2iz2daiJqBCY,118
+tico/experimental/quantization/ptq/wrappers/nn/quant_linear.py,sha256=xW-VEPB7RJoslS3xLVCdhIuMjppknvpkZleRGK4JFVQ,2240
 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
@@ -226,9 +231,9 @@ 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.dev250819.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
-tico-0.1.0.dev250819.dist-info/METADATA,sha256=x4HeKSdCqixV9Xx-eGLbYMAWPvTAGhr0HrMtq0V5lWA,8450
-tico-0.1.0.dev250819.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
-tico-0.1.0.dev250819.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
-tico-0.1.0.dev250819.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
-tico-0.1.0.dev250819.dist-info/RECORD,,
+tico-0.1.0.dev250821.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
+tico-0.1.0.dev250821.dist-info/METADATA,sha256=h3a04PlCzUhqD8MvbgSQHopBPU9IcpNEFx7GthfXVM0,8450
+tico-0.1.0.dev250821.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
+tico-0.1.0.dev250821.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
+tico-0.1.0.dev250821.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
+tico-0.1.0.dev250821.dist-info/RECORD,,