tico 0.1.0.dev250908__py3-none-any.whl → 0.1.0.dev250909__py3-none-any.whl

tico/__init__.py

@@ -29,7 +29,7 @@ __all__ = [
 ]
 
 # THIS LINE IS AUTOMATICALLY GENERATED BY setup.py
-__version__ = "0.1.0.dev250908"
+__version__ = "0.1.0.dev250909"
 
 MINIMUM_SUPPORTED_VERSION = "2.5.0"
 SECURE_TORCH_VERSION = "2.6.0"

tico/experimental/quantization/ptq/wrappers/fairseq/quant_encoder_layer.py

@@ -0,0 +1,165 @@
+# 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.
+#
+# -----------------------------------------------------------------------------
+# This file includes modifications based on fairseq
+#  (https://github.com/facebookresearch/fairseq), originally licensed under
+# the MIT License. See the LICENSE file in the fairseq repository for details.
+# -----------------------------------------------------------------------------
+
+from typing import Optional
+
+import torch.nn as nn
+from torch import Tensor
+
+from tico.experimental.quantization.ptq.quant_config import QuantConfig
+from tico.experimental.quantization.ptq.wrappers.fairseq.quant_mha import (
+    QuantFairseqMultiheadAttention,
+)
+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("fairseq.modules.transformer_layer.TransformerEncoderLayerBase")
+class QuantFairseqEncoderLayer(QuantModuleBase):
+    """
+    Quant-aware drop-in replacement for Fairseq TransformerEncoderLayerBase.
+
+    Design notes (inference-friendly):
+    - All training-time logic (dropout, activation-dropout) is removed.
+    - I/O shape follows Fairseq convention: [T, B, C].
+    - `return_fc` behavior is preserved (returns (x, fc_result) if enabled).
+    """
+
+    def __init__(
+        self,
+        fp_layer: nn.Module,
+        *,
+        qcfg: Optional[QuantConfig] = None,
+        fp_name: Optional[str] = None,
+    ):
+        super().__init__(qcfg, fp_name=fp_name)
+
+        # --- copy meta / config flags from FP layer (read-only) -------------
+        assert hasattr(fp_layer, "embed_dim")
+        assert hasattr(fp_layer, "normalize_before")
+        self.embed_dim: int = int(fp_layer.embed_dim)  # type: ignore[arg-type]
+        self.normalize_before: bool = bool(fp_layer.normalize_before)
+        self.return_fc: bool = bool(getattr(fp_layer, "return_fc", False))
+
+        # --- PTQ-wrapped submodules ----------------------------------------
+        attn_cfg = qcfg.child("self_attn") if qcfg else None
+        fc1_cfg = qcfg.child("fc1") if qcfg else None
+        fc2_cfg = qcfg.child("fc2") if qcfg else None
+        attn_ln_cfg = qcfg.child("self_attn_layer_norm") if qcfg else None
+        final_ln_cfg = qcfg.child("final_layer_norm") if qcfg else None
+
+        assert hasattr(fp_layer, "self_attn") and isinstance(
+            fp_layer.self_attn, nn.Module
+        )
+        assert hasattr(fp_layer, "fc1") and isinstance(fp_layer.fc1, nn.Module)
+        assert hasattr(fp_layer, "fc2") and isinstance(fp_layer.fc2, nn.Module)
+
+        self.self_attn = QuantFairseqMultiheadAttention(
+            fp_layer.self_attn, qcfg=attn_cfg, fp_name=f"{fp_name}.self_attn"
+        )
+        self.fc1 = PTQWrapper(fp_layer.fc1, qcfg=fc1_cfg, fp_name=f"{fp_name}.fc1")
+        self.fc2 = PTQWrapper(fp_layer.fc2, qcfg=fc2_cfg, fp_name=f"{fp_name}.fc2")
+
+        # LayerNorms
+        assert hasattr(fp_layer, "self_attn_layer_norm") and isinstance(
+            fp_layer.self_attn_layer_norm, nn.Module
+        )
+        assert hasattr(fp_layer, "final_layer_norm") and isinstance(
+            fp_layer.final_layer_norm, nn.Module
+        )
+        self.self_attn_layer_norm = PTQWrapper(
+            fp_layer.self_attn_layer_norm,
+            qcfg=attn_ln_cfg,
+            fp_name=f"{fp_name}.self_attn_layer_norm",
+        )
+        self.final_layer_norm = PTQWrapper(
+            fp_layer.final_layer_norm,
+            qcfg=final_ln_cfg,
+            fp_name=f"{fp_name}.final_layer_norm",
+        )
+
+        # Activation function
+        self.activation_fn = fp_layer.activation_fn  # type: ignore[operator]  # e.g., GELU/ReLU
+        self.obs_activation_fn = self._make_obs("activation_fn")
+
+    # ----------------------------------------------------------------------
+    def forward(
+        self,
+        x: Tensor,  # [T,B,C]
+        encoder_padding_mask: Optional[Tensor],
+        attn_mask: Optional[Tensor] = None,  # [T,S] boolean/byte or additive float
+    ):
+        """
+        Returns:
+            x' of shape [T, B, C] (or (x', fc_result) when return_fc=True)
+        """
+        # ---- Self-Attention block (pre-/post-norm kept as in FP layer) ----
+        residual = x
+        if self.normalize_before:
+            x = self.self_attn_layer_norm(x)
+
+        # Fairseq MHA expects [T,B,C]; our wrapped module keeps the same API
+        attn_out, _ = self.self_attn(
+            query=x,
+            key=x,
+            value=x,
+            key_padding_mask=encoder_padding_mask,  # additive float [B,S] or None
+            need_weights=False,
+            attn_mask=attn_mask,  # additive float [T,S] or None
+        )
+        x = residual + attn_out
+
+        if not self.normalize_before:
+            x = self.self_attn_layer_norm(x)
+
+        # ---- FFN block (no dropout/activation-dropout) --------------------
+        residual = x
+        if self.normalize_before:
+            x = self.final_layer_norm(x)
+
+        x = self.fc1(x)  # Linear
+        x = self.activation_fn(x)  # type: ignore[operator]
+        x = self._fq(x, self.obs_activation_fn)
+        x = self.fc2(x)  # Linear
+
+        fc_result = x  # keep before residual for optional return
+
+        x = residual + x
+        if not self.normalize_before:
+            x = self.final_layer_norm(x)
+
+        if self.return_fc:
+            return x, fc_result
+        return x
+
+    def _all_observers(self):
+        yield from (self.obs_activation_fn,)
+        for m in (
+            self.self_attn,
+            self.fc1,
+            self.fc2,
+            self.self_attn_layer_norm,
+            self.final_layer_norm,
+        ):
+            if isinstance(m, QuantModuleBase):
+                yield from m._all_observers()

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

@@ -33,6 +33,7 @@ _CORE_MODULES = (
     "tico.experimental.quantization.ptq.wrappers.llama.quant_decoder_layer",
     "tico.experimental.quantization.ptq.wrappers.llama.quant_mlp",
     # fairseq
+    "tico.experimental.quantization.ptq.wrappers.fairseq.quant_encoder_layer",
     "tico.experimental.quantization.ptq.wrappers.fairseq.quant_mha",
     # add future core wrappers here
 )

{tico-0.1.0.dev250908.dist-info → tico-0.1.0.dev250909.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: tico
-Version: 0.1.0.dev250908
+Version: 0.1.0.dev250909
 Summary: Convert exported Torch module to circle
 Home-page: UNKNOWN
 License: UNKNOWN

{tico-0.1.0.dev250908.dist-info → tico-0.1.0.dev250909.dist-info}/RECORD

@@ -1,4 +1,4 @@
-tico/__init__.py,sha256=KNsLIqvVA_qyf9VOIt_PmHk0L_QUjJqnldhE8_X4n6I,1883
+tico/__init__.py,sha256=c5spmq5DrUrTLuWPal98sdmfFYPjuRym0xcEnK9Am_U,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
@@ -84,8 +84,9 @@ tico/experimental/quantization/ptq/wrappers/__init__.py,sha256=IO6FP_xYbGy0dW0HL
 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=cblE6dpLlK1lh61Xpqxcr2Vwi9XuqdrV5Y7TPSjpETQ,4940
+tico/experimental/quantization/ptq/wrappers/registry.py,sha256=bTd1fZGCXkL4iaduKUXjWVpRXfvOaJGeurxwKJBVu6I,5019
 tico/experimental/quantization/ptq/wrappers/fairseq/__init__.py,sha256=Mc8FLd9DusyB_IT1vk1OYrRkngOYnYd05IvtA9ORVQc,160
+tico/experimental/quantization/ptq/wrappers/fairseq/quant_encoder_layer.py,sha256=aGr80Ku75j2H-UZ0elEa0mOQEyaAs2YJ4WJCN0lonn0,6412
 tico/experimental/quantization/ptq/wrappers/fairseq/quant_mha.py,sha256=HsigmOLeacLXc46QNeFqwQ0DwKQhNrtWTKEtLJoqXoc,15562
 tico/experimental/quantization/ptq/wrappers/llama/__init__.py,sha256=IO6FP_xYbGy0dW0HL26GXD3ouxARaxCK7bz9dn4blPQ,26
 tico/experimental/quantization/ptq/wrappers/llama/quant_attn.py,sha256=-K1COLHIHfJZhQu-RE6KfJIkaL7S6yR4iUj48QkjMTw,8652
@@ -249,9 +250,9 @@ 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.dev250908.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
-tico-0.1.0.dev250908.dist-info/METADATA,sha256=SBFNzxkvYrsWshqSGGkmtVoHtDk1Rp9wVNfhdIRsFdg,8450
-tico-0.1.0.dev250908.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
-tico-0.1.0.dev250908.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
-tico-0.1.0.dev250908.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
-tico-0.1.0.dev250908.dist-info/RECORD,,
+tico-0.1.0.dev250909.dist-info/LICENSE,sha256=kp4JLII7bzRhPb0CPD5XTDZMh22BQ7h3k3B7t8TiSbw,12644
+tico-0.1.0.dev250909.dist-info/METADATA,sha256=udUCPXehe7bxZ27PQKpcxub4Q6VSoaoEmckvpdx5FPo,8450
+tico-0.1.0.dev250909.dist-info/WHEEL,sha256=G16H4A3IeoQmnOrYV4ueZGKSjhipXx8zc8nu9FGlvMA,92
+tico-0.1.0.dev250909.dist-info/entry_points.txt,sha256=kBKYSS_IYrSXmUYevmmepqIVPScq5vF8ulQRu3I_Zf0,59
+tico-0.1.0.dev250909.dist-info/top_level.txt,sha256=oqs7UPoNSKZEwqsX8B-KAWdQwfAa7i60pbxW_Jk7P3w,5
+tico-0.1.0.dev250909.dist-info/RECORD,,