tico 0.1.0.dev250807__py3-none-any.whl → 0.1.0.dev250810__py3-none-any.whl

tico/__init__.py

@@ -29,7 +29,7 @@ __all__ = [
 ]
 
 # THIS LINE IS AUTOMATICALLY GENERATED BY setup.py
-__version__ = "0.1.0.dev250807"
+__version__ = "0.1.0.dev250810"
 
 MINIMUM_SUPPORTED_VERSION = "2.5.0"
 SECURE_TORCH_VERSION = "2.6.0"

tico/experimental/quantization/algorithm/gptq/quantizer.py

@@ -183,7 +183,12 @@ class GPTQQuantizer(BaseQuantizer):
 
         quantizers: Dict[str, Any] = {}
         for l_idx, layer in enumerate(
-            tqdm(target_layers, desc="Quantizing layers", unit="layer")
+            tqdm(
+                target_layers,
+                desc="Quantizing layers",
+                unit="layer",
+                disable=not gptq_conf.show_progress,
+            )
         ):
             # 1) Identify quantizable submodules within the layer
             full = find_layers(layer)
@@ -218,6 +223,7 @@ class GPTQQuantizer(BaseQuantizer):
                     desc=f"[L{l_idx}] collecting",
                     leave=False,
                     unit="batch",
+                    disable=not gptq_conf.show_progress,
                 ):
                     cache_args_batch = gather_single_batch_from_list(
                         self.cache_args, batch_idx
@@ -251,6 +257,7 @@ class GPTQQuantizer(BaseQuantizer):
                 desc=f"[L{l_idx}] re-forward",
                 leave=False,
                 unit="batch",
+                disable=not gptq_conf.show_progress,
             ):
                 cache_args_batch = gather_single_batch_from_list(
                     self.cache_args, batch_idx

tico/experimental/quantization/config.py

@@ -42,8 +42,9 @@ class GPTQConfig(BaseConfig):
     Configuration for GPTQ.
     """
 
-    def __init__(self, verbose: bool = False):
+    def __init__(self, verbose: bool = False, show_progress: bool = True):
         self.verbose = verbose
+        self.show_progress = show_progress
 
     @property
     def name(self) -> str:

tico/utils/dtype.py

@@ -1,6 +1,8 @@
 import numpy as np
 import torch
 
+from circle_schema import circle
+
 NUMPY_TO_TORCH_DTYPE_DICT = {
     np.dtype("float32"): torch.float32,
     np.dtype("float64"): torch.float64,
@@ -15,6 +17,26 @@ NUMPY_TO_TORCH_DTYPE_DICT = {
     np.dtype("bool"): torch.bool,
 }
 
+CIRCLE_TO_TORCH_DTYPE_DICT = {
+    circle.TensorType.TensorType.FLOAT32: torch.float32,
+    circle.TensorType.TensorType.UINT8: torch.uint8,
+    circle.TensorType.TensorType.INT8: torch.int8,
+    circle.TensorType.TensorType.INT16: torch.int16,
+    circle.TensorType.TensorType.INT32: torch.int32,
+    circle.TensorType.TensorType.INT64: torch.int64,
+    circle.TensorType.TensorType.BOOL: torch.bool,
+}
+
 
 def numpy_dtype_to_torch_dtype(np_dtype: np.dtype) -> torch.dtype:
     return NUMPY_TO_TORCH_DTYPE_DICT[np_dtype]
+
+
+def circle_dtype_to_torch_dtype(circle_dtype: int) -> torch.dtype:
+    assert isinstance(circle_dtype, int)
+    if circle_dtype not in CIRCLE_TO_TORCH_DTYPE_DICT:
+        raise RuntimeError(f"Unsupported dtype {circle_dtype}")
+
+    torch_dtype = CIRCLE_TO_TORCH_DTYPE_DICT[circle_dtype]
+    assert torch_dtype is not None
+    return torch_dtype

tico/utils/signature.py

@@ -0,0 +1,248 @@
+# 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 Sequence
+
+import numpy as np
+import torch
+from circle_schema import circle
+
+from tico.serialize.circle_mapping import to_circle_shape
+from tico.utils.dtype import circle_dtype_to_torch_dtype
+from tico.utils.installed_packages import is_dynamic_cache_available
+
+
+def is_dynamic_cache_instance(value):
+    if is_dynamic_cache_available():
+        from transformers.cache_utils import DynamicCache
+
+        return isinstance(value, DynamicCache)
+    else:
+        return False
+
+
+def flatten_and_convert_kwargs(kwargs: dict) -> dict[str, torch.Tensor]:
+    result = {}  # type: ignore[var-annotated]
+    for k, v in kwargs.items():
+        if v is None:
+            continue
+        elif isinstance(v, (list, tuple)):
+            # 1. handle list
+            def unpack_recursive(name, value, store=None):
+                if store is None:
+                    store = {}
+
+                if isinstance(value, (tuple, list)):
+                    for i, v in enumerate(value):
+                        # recursive call. Append index to name and explore lower level
+                        unpack_recursive(f"{name}_{i}", v, store)
+                else:
+                    # base type (scalar etc.) directly stored
+                    store[name] = value
+
+                return store
+
+            unpack_recursive(k, v, result)
+        elif is_dynamic_cache_instance(v):
+            # 2. handle DynamicCache
+            for idx, cache_val in enumerate(v.key_cache):
+                result[f"{k}_key_cache_{idx}"] = cache_val
+
+            for idx, cache_val in enumerate(v.value_cache):
+                result[f"{k}_value_cache_{idx}"] = cache_val
+        else:
+            result[k] = v
+
+    # 3. Convert to tensors
+    for k, v in result.items():
+        result[k] = v if isinstance(v, torch.Tensor) else torch.tensor(v)
+
+    return result
+
+
+def flatten_and_convert_args(args: Sequence) -> tuple:
+    result = []  # type: ignore[var-annotated]
+    for item in args:
+        if item is None:
+            continue
+
+        # 1. recursion on list and tuple
+        if isinstance(item, (list, tuple)):
+            result.extend(flatten_and_convert_args(item))
+            continue
+
+        # 2. handle DynamicCache
+        if is_dynamic_cache_available():
+            from transformers.cache_utils import DynamicCache
+
+            if isinstance(item, DynamicCache):
+                # NOTE The tensor order is: key_in → key_out → value_in → value_out
+                #
+                # Refer to https://github.com/huggingface/transformers/blob/3457e8e73e4f5532cc69059682b1ba4484d7e7e8/src/transformers/cache_utils.py#L557
+                # ```
+                # self.key_cache[layer_idx] = torch.cat([self.key_cache[layer_idx], key_states], dim=-2)
+                # self.value_cache[layer_idx] = torch.cat([self.value_cache[layer_idx], value_states], dim=-2)
+                # ```
+                result.extend(item.key_cache)
+                result.extend(item.value_cache)
+                continue
+
+        # 3. Convert to tensors
+        result.append(item if isinstance(item, torch.Tensor) else torch.tensor(item))
+
+    return tuple(result)
+
+
+class ModelInputSpec:
+    @classmethod
+    def load(cls, circle_path):
+        def load(circle_path: str) -> bytes:
+            with open(circle_path, "rb") as f:
+                buf = bytes(f.read())
+            return buf
+
+        circle_binary = load(circle_path)
+        return cls(circle_binary)
+
+    def __init__(self, circle_binary):
+        model = circle.Model.Model.GetRootAsModel(circle_binary, 0)
+        assert model.SubgraphsLength() == 1, "Only one subgraph is supported"
+
+        graph = model.Subgraphs(0)
+        tensors = [graph.Tensors(graph.Inputs(o)) for o in range(graph.InputsLength())]
+
+        self.names = [t.Name().decode("utf-8").split("::")[-1] for t in tensors]
+        self.shapes = [t.ShapeAsNumpy() for t in tensors]
+        self.shape_signatures = list(
+            map(
+                lambda x: None if (isinstance(x, int) and x == 0) else x,
+                (t.ShapeSignatureAsNumpy() for t in tensors),
+            )
+        )
+        self.types: list[torch.dtype] = [
+            circle_dtype_to_torch_dtype(t.Type()) for t in tensors
+        ]
+        self.name_to_idx = {name: idx for idx, name in enumerate(self.names)}
+
+    def bind(self, args, kwargs, check=True):
+        """Convert args and kwargs into an ordered list according to model input order"""
+        inputs = []
+        args = flatten_and_convert_args(args)
+        kwargs = flatten_and_convert_kwargs(kwargs)
+
+        # 1. positional arguments
+        for i, val in enumerate(args):
+            if i >= len(self.names):
+                raise ValueError(f"Too many positional arguments ({i+1}).")
+            name = self.names[i]
+            if name in kwargs:
+                raise TypeError(
+                    f"Got multiple values for argument '{name}' (positional and keyword)."
+                )
+            inputs.append(val)
+
+        # 2. keyword arguments
+        for idx in range(len(args), len(self.names)):
+            name = self.names[idx]
+            if name not in kwargs:
+                raise ValueError(f"Missing argument for input '{name}'.")
+            inputs.append(kwargs[name])
+
+        if check:
+            self.check_types(inputs)
+            self.check_shapes(inputs)
+
+        return inputs
+
+    def check_types(self, inputs):
+        """Check the types of input values"""
+        for i, (inp, ref_type) in enumerate(zip(inputs, self.types)):
+            # TODO: Support more data types (np array)
+            assert isinstance(
+                inp, (torch.Tensor | int | float)
+            ), f"Input '{self.names[i]}' type must be a torch tensor or scalar."
+
+            if isinstance(inp, torch.Tensor):
+                if inp.dtype != ref_type:
+                    raise TypeError(
+                        f"Input '{self.names[i]}' type {inp.dtype} != expected {ref_type}"
+                    )
+            else:
+                # Scalars (int, float)
+                if ref_type == torch.float32:
+                    if not isinstance(inp, (float)):
+                        raise TypeError(
+                            f"Input '{self.names[i]}' type {type(inp)} != expected {ref_type}"
+                        )
+                elif ref_type == torch.int64:
+                    if not isinstance(inp, (int)):
+                        raise TypeError(
+                            f"Input '{self.names[i]}' type {type(inp)} != expected {ref_type}"
+                        )
+                else:
+                    print(f"Unexpected ref_type: {ref_type}")
+
+    def check_shapes(self, inputs):
+        """Check the shapes of input values"""
+
+        def merge(shape, shape_sig):
+            """
+            Merge shape signature with shape
+            """
+            from copy import deepcopy
+
+            shape_merged = deepcopy(shape)
+            if shape_sig is not None:
+                for idx, ss in enumerate(shape_sig):
+                    if ss == -1:
+                        shape_merged[idx] = -1
+
+            return shape_merged
+
+        for i, (inp, ref_shape, ref_shape_sig) in enumerate(
+            zip(inputs, self.shapes, self.shape_signatures)
+        ):
+            # TODO: Support more data types (np array)
+            assert isinstance(
+                inp, (torch.Tensor | int | float)
+            ), f"Input '{self.names[i]}' type must be a torch tensor or scalar."
+
+            if isinstance(inp, torch.Tensor):  # Tensor
+                in_shape, in_shape_sig = to_circle_shape(inp.size())
+
+                if len(in_shape) != len(ref_shape):
+                    raise ValueError(
+                        f"Input '{self.names[i]}' has invalid rank {len(in_shape)}!= expected {len(ref_shape)}"
+                    )
+
+                in_merged_shape = merge(in_shape, in_shape_sig)
+                ref_merged_shape = merge(ref_shape, ref_shape_sig)
+                for in_shp, ref_shp in zip(in_merged_shape, ref_merged_shape):
+                    if ref_shp == -1:
+                        continue
+                    if in_shp == -1:
+                        raise ValueError(
+                            f"Input '{self.names[i]}' has unknown dimension {inp.size()} != expected shape({ref_shape}) / shape signature({ref_shape_sig}) "
+                        )
+                    if in_shp != ref_shp:
+                        raise ValueError(
+                            f"Input '{self.names[i]}' has wrong dimension {inp.size()} != expected shape({ref_shape}) / shape signature({ref_shape_sig}) "
+                        )
+            elif isinstance(inp, (int, float)):  # Scalar
+                if len(ref_shape) > 0:
+                    raise ValueError(
+                        f"Input '{self.names[i]}' has invalid rank {len(ref_shape)}"
+                    )
+            else:
+                print(f"Unexpected input type: {type(inp)}")

{tico-0.1.0.dev250807.dist-info → tico-0.1.0.dev250810.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: tico
-Version: 0.1.0.dev250807
+Version: 0.1.0.dev250810
 Summary: Convert exported Torch module to circle
 Home-page: UNKNOWN
 License: UNKNOWN

{tico-0.1.0.dev250807.dist-info → tico-0.1.0.dev250810.dist-info}/RECORD

@@ -1,4 +1,4 @@
-tico/__init__.py,sha256=xXynWlns6HYdl6Zf9xwJtdTnFkscm8b49vqhn8FYJZg,1883
+tico/__init__.py,sha256=H67V6wHv1jk6smTy2N_nzF9vFR0U8Rf-Vsc19oWjhpM,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
@@ -6,14 +6,14 @@ tico/config/factory.py,sha256=il0zqB6Lm5NX2LnG-TUhmiP9vVeZ_3TucJMorVZIodY,1324
 tico/config/v1.py,sha256=O1jzpUBDwoWpLohEpI08pJNwVB-yz3ufPrQm2_XWq4Y,1108
 tico/experimental/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 tico/experimental/quantization/__init__.py,sha256=IaJPZegVJp0P3luutBo907Kp5sOJensE1Mm-XBG_jBs,122
-tico/experimental/quantization/config.py,sha256=h01WpP8Y-dLj6yg12pMZm3PXJqUnU2sWip5jBRc5x9Q,1604
+tico/experimental/quantization/config.py,sha256=1bCSAUI043Kbq08j59mb-K1cP2lmBMbekh8p3hNK6b8,1675
 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/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=5Sai9gD__6mt7rrUJutJ2L-BV_2biUXcvZC1H-a16Vc,11467
+tico/experimental/quantization/algorithm/gptq/quantizer.py,sha256=_ZnSD_LBag_FVcVEniPKBmw7bNZ2iZLZ8aZnexnCgrs,11693
 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
@@ -190,7 +190,7 @@ tico/utils/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 tico/utils/convert.py,sha256=GgZwZtiqFzTdszfUQO0vcX39lKjs97gYwZ-Tiw_4Bbo,13222
 tico/utils/define.py,sha256=Ypgp7YffM4pgPl4Zh6TmogSn1OxGBMRw_e09qYGflZk,1467
 tico/utils/diff_graph.py,sha256=_eDGGPDPYQD4b--MXX0DLoVgSt_wLfNPt47UlolLLR4,5272
-tico/utils/dtype.py,sha256=4-k1iUaHivFFXAQuDs7up6fXt5y4FqldGNokAPa3kic,603
+tico/utils/dtype.py,sha256=L5Qb7qgbt0eQ5frUTvHYrRtTJb1dg4-JNEopcxCNg1U,1389
 tico/utils/errors.py,sha256=f3csJjgbXG9W1aHhqEcou008Aor19W57X8oT5Hx8w1M,954
 tico/utils/graph.py,sha256=jD6m58m5JmN9mPfaROA9CW3406iJxmnukke00AuwRqI,9131
 tico/utils/installed_packages.py,sha256=J0FTwnkCGs0MxRWoCMYAqiwH7Z0GWFDLV--x-IndSp4,1017
@@ -202,6 +202,7 @@ tico/utils/pytree_utils.py,sha256=jrk3N6X6LiUnBCX_gM1K9nywbVAJBVnszlTAgeIeDUc,52
 tico/utils/record_input.py,sha256=QN-8D71G_WAX3QQQ5CIwbEfFJZTQ3CvL4wCMiVddua4,3894
 tico/utils/register_custom_op.py,sha256=3-Yl6iYmx1qQA2igNHt4hYhQhQMkdPb7gF50LIY8yvc,27350
 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
@@ -210,9 +211,9 @@ 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.dev250807.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
-tico-0.1.0.dev250807.dist-info/METADATA,sha256=ATnzGR44gSRsXtSTjRBwY0DOQK512GtszXrNo3r3rD4,8450
-tico-0.1.0.dev250807.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
-tico-0.1.0.dev250807.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
-tico-0.1.0.dev250807.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
-tico-0.1.0.dev250807.dist-info/RECORD,,
+tico-0.1.0.dev250810.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
+tico-0.1.0.dev250810.dist-info/METADATA,sha256=N1kcg1vk8kn6bLJgRUSZMZalY-mTn46jqLvVj9NGvR4,8450
+tico-0.1.0.dev250810.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
+tico-0.1.0.dev250810.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
+tico-0.1.0.dev250810.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
+tico-0.1.0.dev250810.dist-info/RECORD,,