tico 0.1.0.dev250825__py3-none-any.whl → 0.1.0.dev250827__py3-none-any.whl

tico/__init__.py

@@ -29,7 +29,7 @@ __all__ = [
 ]
 
 # THIS LINE IS AUTOMATICALLY GENERATED BY setup.py
-__version__ = "0.1.0.dev250825"
+__version__ = "0.1.0.dev250827"
 
 MINIMUM_SUPPORTED_VERSION = "2.5.0"
 SECURE_TORCH_VERSION = "2.6.0"

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

@@ -0,0 +1 @@
+# DO NOT REMOVE THIS FILE

tico/experimental/quantization/ptq/examples/quantize_linear.py

@@ -0,0 +1,106 @@
+# 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.
+
+# =============================================================================
+# POST-TRAINING QUANTIZATION EXAMPLE — Simple Linear Model
+# -----------------------------------------------------------------------------
+# This demo shows a minimal PTQ flow for a toy model:
+#   1. Define a simple model with a single Linear layer.
+#   2. Replace the FP32 Linear with a QuantLinear wrapper.
+#   3. Run a short calibration pass to collect activation statistics.
+#   4. Freeze scales / zero-points and switch to INT-simulation mode.
+#   5. Compare INT vs FP32 outputs with a mean-absolute-diff check.
+#   6. Export the quantized model to a Circle format.
+# =============================================================================
+
+import pathlib
+
+import torch
+import torch.nn as nn
+
+from tico.experimental.quantization.evaluation.metric import compute_peir
+from tico.experimental.quantization.evaluation.utils import plot_two_outputs
+
+from tico.experimental.quantization.ptq.mode import Mode
+from tico.experimental.quantization.ptq.wrappers.nn.quant_linear import QuantLinear
+from tico.utils.utils import SuppressWarning
+
+# -------------------------------------------------------------------------
+# 0. Define a toy model (1 Linear layer only)
+# -------------------------------------------------------------------------
+class TinyLinearModel(nn.Module):
+    """A minimal model: single Linear layer."""
+
+    def __init__(self, in_features=16, out_features=8):
+        super().__init__()
+        self.fc = nn.Linear(in_features, out_features, bias=False)
+
+    def forward(self, x):
+        return self.fc(x)
+
+
+# Instantiate FP32 model
+model = TinyLinearModel()
+model.eval()
+
+# Keep FP32 reference for diff check
+fp32_layer = model.fc
+
+# -------------------------------------------------------------------------
+# 1. Replace the Linear with QuantLinear wrapper
+# -------------------------------------------------------------------------
+model.fc = QuantLinear(fp32_layer)  # type: ignore[assignment]
+# model.fc = PTQWrapper(fp32_layer) (Wrapping helper class)
+qlayer = model.fc  # alias for brevity
+
+# -------------------------------------------------------------------------
+# 2. Single-pass calibration (collect activation ranges)
+# -------------------------------------------------------------------------
+assert isinstance(qlayer, QuantLinear)
+with torch.no_grad():
+    qlayer.enable_calibration()
+    for _ in range(16):  # small toy batch
+        x = torch.randn(4, 16)  # (batch=4, features=16)
+        _ = model(x)
+    qlayer.freeze_qparams()  # lock scales & zero-points
+
+assert qlayer._mode is Mode.QUANT, "Quantization mode should be active now."
+
+# -------------------------------------------------------------------------
+# 3. Quick INT-sim vs FP32 sanity check
+# -------------------------------------------------------------------------
+x = torch.randn(2, 16)
+with torch.no_grad():
+    int8_out = model(x)
+    fp32_out = fp32_layer(x)
+
+print("┌───────────── Quantization Error Summary ─────────────")
+print(f"│ Mean |diff|: {(int8_out - fp32_out).abs().mean().item():.6f}")
+print(f"│ PEIR       : {compute_peir(fp32_out, int8_out) * 100:.6f} %")
+print("└──────────────────────────────────────────────────────")
+print(plot_two_outputs(fp32_out, int8_out))
+
+# -------------------------------------------------------------------------
+# 4. Export the calibrated model to Circle
+# -------------------------------------------------------------------------
+import tico
+
+save_path = pathlib.Path("tiny_linear.q.circle")
+example_input = torch.randn(1, 16)
+
+with SuppressWarning(UserWarning, ".*"):
+    cm = tico.convert(model, (example_input,))  # forward(x) only
+cm.save(save_path)
+
+print(f"Quantized Circle model saved to {save_path.resolve()}")

tico/experimental/quantization/ptq/examples/quantize_llama_mlp.py

@@ -0,0 +1,87 @@
+# 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 pathlib
+
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+import tico
+from tico.experimental.quantization.evaluation.metric import compute_peir
+from tico.experimental.quantization.evaluation.utils import plot_two_outputs
+from tico.experimental.quantization.ptq.dtypes import INT16
+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.llama.quant_mlp import QuantLlamaMLP
+from tico.utils.utils import SuppressWarning
+
+name = "Maykeye/TinyLLama-v0"
+model = AutoModelForCausalLM.from_pretrained(name)
+tokenizer = AutoTokenizer.from_pretrained(name)
+model.eval()
+
+# -------------------------------------------------------------------------
+# 1. Replace layer-0’s MLP with QuantLlamaMLP
+# -------------------------------------------------------------------------
+fp32_mlp = model.model.layers[0].mlp
+model.model.layers[0].mlp = QuantLlamaMLP(
+    fp32_mlp,
+    qcfg=QuantConfig(default_dtype=INT16, default_qscheme=QScheme.PER_TENSOR_SYMM),
+)  # PTQWrapper(fp32_mlp) is also fine
+model.eval()
+
+mlp_q = model.model.layers[0].mlp
+
+# -------------------------------------------------------------------------
+# 2. Single-pass calibration
+# -------------------------------------------------------------------------
+with torch.no_grad():
+    mlp_q.enable_calibration()
+    for _ in range(16):
+        prompts = ["hello tinyllama "] * 8
+        enc = tokenizer(prompts, return_tensors="pt")
+        emb = model.model.embed_tokens(enc["input_ids"])
+        _ = mlp_q(emb)
+
+    mlp_q.freeze_qparams()
+
+assert mlp_q._mode is Mode.QUANT, "Quantization mode should be active now."
+
+# -------------------------------------------------------------------------
+# 3. Quick diff check (INT-sim vs FP32)
+# -------------------------------------------------------------------------
+with torch.no_grad():
+    ids = tokenizer("quant all tensors!", return_tensors="pt")
+    emb = model.model.embed_tokens(ids["input_ids"])
+    int16 = mlp_q(emb)  # INT-sim
+    fp32 = fp32_mlp(emb)  # baseline reference
+
+print("┌───────────── Quantization Error Summary ─────────────")
+print(f"│ Mean |diff|: {(int16 - fp32).abs().mean().item():.6f}")
+print(f"│ PEIR       : {compute_peir(fp32, int16) * 100:.6f} %")
+print("└──────────────────────────────────────────────────────")
+print(plot_two_outputs(fp32, int16))
+
+# -------------------------------------------------------------------------
+# 4. Export the quantized block
+# -------------------------------------------------------------------------
+save_path = pathlib.Path("mlp.q.circle")
+example_in = (torch.randn(1, 1, model.config.hidden_size),)
+
+with SuppressWarning(UserWarning, ".*"):
+    cm = tico.convert(mlp_q, example_in)
+cm.save(save_path)
+
+print(f"Quantized Circle model saved to {save_path.resolve()}")

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

@@ -0,0 +1,3 @@
+from tico.experimental.quantization.ptq.wrappers.llama.quant_mlp import QuantLlamaMLP
+
+__all__ = ["QuantLlamaMLP"]

tico/experimental/quantization/ptq/wrappers/llama/quant_mlp.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 typing import Optional
+
+import torch
+import torch.nn as nn
+
+from tico.experimental.quantization.ptq.quant_config import QuantConfig
+from tico.experimental.quantization.ptq.wrappers.ptq_wrapper import PTQWrapper
+from tico.experimental.quantization.ptq.wrappers.quant_module_base import (
+    QuantModuleBase,
+)
+from tico.experimental.quantization.ptq.wrappers.registry import try_register
+
+
+@try_register("transformers.models.llama.modeling_llama.LlamaMLP")
+class QuantLlamaMLP(QuantModuleBase):
+    def __init__(
+        self,
+        mlp_fp: nn.Module,
+        *,
+        qcfg: Optional[QuantConfig] = None,
+        fp_name: Optional[str] = None,
+    ):
+        super().__init__(qcfg, fp_name=fp_name)
+
+        # ----- child configs (hierarchical override) -------------------
+        gate_cfg = qcfg.child("gate_proj") if qcfg else None
+        up_cfg = qcfg.child("up_proj") if qcfg else None
+        down_cfg = qcfg.child("down_proj") if qcfg else None
+        act_cfg = qcfg.child("act_fn") if qcfg else None
+
+        # ----- wrap three Linear layers -------------------------------
+        assert hasattr(mlp_fp, "gate_proj") and isinstance(
+            mlp_fp.gate_proj, torch.nn.Module
+        )
+        assert hasattr(mlp_fp, "up_proj") and isinstance(
+            mlp_fp.up_proj, torch.nn.Module
+        )
+        assert hasattr(mlp_fp, "down_proj") and isinstance(
+            mlp_fp.down_proj, torch.nn.Module
+        )
+        self.gate_proj = PTQWrapper(
+            mlp_fp.gate_proj, qcfg=gate_cfg, fp_name=f"{fp_name}.gate_proj"
+        )
+        self.up_proj = PTQWrapper(
+            mlp_fp.up_proj, qcfg=up_cfg, fp_name=f"{fp_name}.up_proj"
+        )
+        self.down_proj = PTQWrapper(
+            mlp_fp.down_proj, qcfg=down_cfg, fp_name=f"{fp_name}.down_proj"
+        )
+
+        # ----- activation ---------------------------------------------
+        assert hasattr(mlp_fp, "act_fn") and isinstance(mlp_fp.act_fn, torch.nn.Module)
+        self.act_fn = PTQWrapper(
+            mlp_fp.act_fn, qcfg=act_cfg, fp_name=f"{fp_name}.act_fn"
+        )
+
+        # ----- local observers ----------------------------------------
+        self.act_in_obs = self._make_obs("act_in")
+        self.mul_obs = self._make_obs("mul")
+
+    def forward(self, x: torch.Tensor):
+        # 1) quantize input once
+        x_q = self._fq(x, self.act_in_obs)
+
+        # 2) parallel projections
+        g = self.gate_proj(x_q)
+        u = self.up_proj(x_q)
+
+        # 3) activation on gate
+        a = self.act_fn(g)
+
+        # 4) element-wise product
+        h = self._fq(a * u, self.mul_obs)
+
+        # 5) final projection
+        return self.down_proj(h)
+
+    def _all_observers(self):
+        # local first
+        yield self.act_in_obs
+        yield self.mul_obs
+        # recurse into children that are QuantModuleBase
+        for m in (self.gate_proj, self.up_proj, self.down_proj, self.act_fn):
+            yield from m._all_observers()

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

@@ -48,7 +48,24 @@ class PTQWrapper(QuantModuleBase):
         return self.wrapped(*args, **kwargs)
 
     def _all_observers(self):
-        yield from self.wrapped._all_observers()
+        """
+        PTQWrapper itself owns NO observers (transparent node).
+        Returning an empty iterator prevents double-processing when parents
+        traverse the tree and then recurse into `self.wrapped`.
+        """
+        return ()  # no local observers
+
+    def named_observers(self):
+        """
+        Proxy to the wrapped module so debugging tools can still enumerate observers.
+        """
+        yield from self.wrapped.named_observers()
+
+    def get_observer(self, name: str):
+        """
+        Proxy to the wrapped module for direct lookup by name.
+        """
+        return self.wrapped.get_observer(name)
 
     def extra_repr(self) -> str:
         return self.wrapped.extra_repr()

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

@@ -28,6 +28,7 @@ _CORE_MODULES = (
     "tico.experimental.quantization.ptq.wrappers.nn.quant_layernorm",
     "tico.experimental.quantization.ptq.wrappers.nn.quant_linear",
     "tico.experimental.quantization.ptq.wrappers.nn.quant_silu",
+    "tico.experimental.quantization.ptq.wrappers.llama.quant_mlp",
     # add future core wrappers here
 )
 

{tico-0.1.0.dev250825.dist-info → tico-0.1.0.dev250827.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: tico
-Version: 0.1.0.dev250825
+Version: 0.1.0.dev250827
 Summary: Convert exported Torch module to circle
 Home-page: UNKNOWN
 License: UNKNOWN

{tico-0.1.0.dev250825.dist-info → tico-0.1.0.dev250827.dist-info}/RECORD

@@ -1,4 +1,4 @@
-tico/__init__.py,sha256=qQ6AhQGEdqTIol9a3pZd8krL41W7mwvr0HoA1tPCeqU,1883
+tico/__init__.py,sha256=zW-Qvhbt2u7nt1XfXlgpi6kJrCNuyIJ3PddICgrWpZI,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
@@ -61,6 +61,9 @@ tico/experimental/quantization/ptq/dtypes.py,sha256=xfCBtq6mQmUYRwsoFgII6gvRl1ra
 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/quant_config.py,sha256=nm7570Y1X2mOT_8s27ilWid04otor6cVTi9GwgAEaKc,4300
+tico/experimental/quantization/ptq/examples/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
+tico/experimental/quantization/ptq/examples/quantize_linear.py,sha256=8zq-ZJDYgam0xQ-PbC6Xb1I7W1mv0Wi-b--IP2wwXtw,4539
+tico/experimental/quantization/ptq/examples/quantize_llama_mlp.py,sha256=MktXLIxz2fUAA7sDG_wN6B8eUi71jOgAaMfhPj9u9XM,3736
 tico/experimental/quantization/ptq/observers/__init__.py,sha256=WF2MvL9M_jl-B1FqcY9zic34NOCRp17HkRYv-TMxMr4,613
 tico/experimental/quantization/ptq/observers/affine_base.py,sha256=e2Eba64nrxKQyE4F_WJ7WTSsk3xe6bkdGUKaoLFWGFw,4638
 tico/experimental/quantization/ptq/observers/base.py,sha256=Wons1MzpqK1mfcy-ppl-B2Dum0edXg2dWW2Lw3V18tw,3280
@@ -71,10 +74,12 @@ 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/ptq_wrapper.py,sha256=F9sK_DiRaXiGNHULcwIbs5EUtHz6ZJ7N4r5CWTTfhsM,2442
 tico/experimental/quantization/ptq/wrappers/quant_elementwise.py,sha256=LhEoobfvto6zKrBOKL4gmxfFFc31jHzyQV_zfps-iQM,3604
 tico/experimental/quantization/ptq/wrappers/quant_module_base.py,sha256=vkcDos_knGSS29rIZuEIWkAJLHrENbGz8nCH2-iara8,5969
-tico/experimental/quantization/ptq/wrappers/registry.py,sha256=562nKSlp9qF-w4-aQeJbx2V_wMGE2FRrjIKUfRwC4Mg,4571
+tico/experimental/quantization/ptq/wrappers/registry.py,sha256=utZY381tuKYyUITptUfx4_0V-jprSjvEXpGtfa4cPnI,4638
+tico/experimental/quantization/ptq/wrappers/llama/__init__.py,sha256=vVdVj7HMjxRPAYEO4DEg1l_7qztMmGCGCe7GbniCzrM,115
+tico/experimental/quantization/ptq/wrappers/llama/quant_mlp.py,sha256=uZMnrX66oZwxhKhcNbLXXeri-WxxRBiZnr15aBXJMm0,3562
 tico/experimental/quantization/ptq/wrappers/nn/__init__.py,sha256=I9uTt5HfcRoMEDYHpAeATMv2TbCQiX0ZbfUFMzSJ4Qw,336
 tico/experimental/quantization/ptq/wrappers/nn/quant_layernorm.py,sha256=G5Sgt-tXnzh0Rxyk-2honmZIfEQOZlRfOsoDBdSGmA4,6887
 tico/experimental/quantization/ptq/wrappers/nn/quant_linear.py,sha256=xW-VEPB7RJoslS3xLVCdhIuMjppknvpkZleRGK4JFVQ,2240
@@ -233,9 +238,9 @@ 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.dev250825.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
-tico-0.1.0.dev250825.dist-info/METADATA,sha256=7wBkNIwJG_prscPdY7Rn_Muit4OuPN29Q8C_isHlEdI,8450
-tico-0.1.0.dev250825.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
-tico-0.1.0.dev250825.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
-tico-0.1.0.dev250825.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
-tico-0.1.0.dev250825.dist-info/RECORD,,
+tico-0.1.0.dev250827.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
+tico-0.1.0.dev250827.dist-info/METADATA,sha256=dkNactHh3C0pzY9_mblMYxzAdqWIRbFTre9sBiKfDAo,8450
+tico-0.1.0.dev250827.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
+tico-0.1.0.dev250827.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
+tico-0.1.0.dev250827.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
+tico-0.1.0.dev250827.dist-info/RECORD,,