tico 0.1.0.dev250904__py3-none-any.whl → 0.1.0.dev251109__py3-none-any.whl

tico/quantization/wrapq/wrappers/fairseq/quant_encoder.py

@@ -0,0 +1,331 @@
+# 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.
+# -----------------------------------------------------------------------------
+
+import math
+from typing import Dict, List, Literal, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from torch import Tensor
+
+from tico.quantization.config.ptq import PTQConfig
+from tico.quantization.wrapq.wrappers.ptq_wrapper import PTQWrapper
+from tico.quantization.wrapq.wrappers.quant_module_base import QuantModuleBase
+from tico.quantization.wrapq.wrappers.registry import try_register
+
+
+@try_register("fairseq.models.transformer.TransformerEncoderBase")
+class QuantFairseqEncoder(QuantModuleBase):
+    """
+    Quant-aware drop-in replacement for Fairseq TransformerEncoderBase.
+
+    Key design choices:
+    - Keep embeddings and LayerNorms in FP.
+    - Remove training-time logic (dropout, activation-dropout, quant_noise).
+    - Attention masks are handled statically inside the layer wrapper; this
+      encoder only does the original padding zero-out before the stack.
+
+    I/O contracts:
+    - Forward signature and returned dictionary are identical to the original
+      when `use_external_inputs=False`.
+    - When `use_external_inputs=True`, forward returns a single Tensor (T,B,C)
+      and completely skips embedding/positional/LN/mask-creation paths.
+    - Tensor shapes follow Fairseq convention.
+    """
+
+    def __init__(
+        self,
+        fp_encoder: nn.Module,
+        *,
+        qcfg: Optional[PTQConfig] = None,
+        fp_name: Optional[str] = None,
+        use_external_inputs: bool = False,  # export-mode flag
+        return_type: Literal["tensor", "dict"] = "dict",
+    ):
+        super().__init__(qcfg, fp_name=fp_name)
+        self.use_external_inputs = use_external_inputs
+        self.return_type: Literal["tensor", "dict"] = return_type
+
+        # --- carry basic config / metadata (read-only copies) ---------------
+        assert hasattr(fp_encoder, "cfg")
+        self.cfg = fp_encoder.cfg
+        self.return_fc: bool = bool(getattr(fp_encoder, "return_fc", False))
+
+        # Embedding stack ----------------------------------------------------
+        assert hasattr(fp_encoder, "embed_tokens") and isinstance(
+            fp_encoder.embed_tokens, nn.Module
+        )
+        self.embed_tokens = fp_encoder.embed_tokens  # keep FP embeddings
+
+        assert hasattr(fp_encoder, "padding_idx")
+        self.padding_idx: int = int(fp_encoder.padding_idx)  # type: ignore[arg-type]
+
+        # scale = sqrt(embed_dim) unless disabled
+        embed_dim = int(self.embed_tokens.embedding_dim)  # type: ignore[arg-type]
+        no_scale = bool(getattr(self.cfg, "no_scale_embedding", False))
+        self.embed_scale: float = 1.0 if no_scale else math.sqrt(embed_dim)
+
+        # Positional embeddings (keep as-is; no FQ)
+        self.embed_positions = getattr(fp_encoder, "embed_positions", None)
+        # Optional embedding LayerNorm
+        self.layernorm_embedding = getattr(fp_encoder, "layernorm_embedding", None)
+
+        # Final encoder LayerNorm (pre-norm stacks may set this to None)
+        self.layer_norm = getattr(fp_encoder, "layer_norm", None)
+
+        # Max positions (reuse for API parity)
+        self.max_source_positions: int = int(fp_encoder.max_source_positions)  # type: ignore[arg-type]
+
+        # --- wrap encoder layers with PTQWrapper ----------------------------
+        assert hasattr(fp_encoder, "layers")
+        fp_layers = list(fp_encoder.layers)  # type: ignore[arg-type]
+        self.layers = nn.ModuleList()
+
+        # Prepare child PTQConfig namespaces: layers/<idx>
+        layers_qcfg = qcfg.child("layers") if qcfg else None
+        for i, layer in enumerate(fp_layers):
+            child_cfg = layers_qcfg.child(str(i)) if layers_qcfg else None
+            self.layers.append(
+                PTQWrapper(layer, qcfg=child_cfg, fp_name=f"{fp_name}.layers.{i}")
+            )
+
+        # Version buffer (keep for state_dict parity)
+        version = getattr(fp_encoder, "version", None)
+        if isinstance(version, torch.Tensor):
+            self.register_buffer("version", version.clone(), persistent=False)
+        else:
+            self.register_buffer("version", torch.tensor([3.0]), persistent=False)
+
+    # ----------------------------------------------------------------------
+    def forward_embedding(
+        self, src_tokens: Tensor, token_embedding: Optional[Tensor] = None
+    ) -> Tuple[Tensor, Tensor]:
+        """
+        Embed tokens and add positional embeddings. Dropout/quant_noise are removed.
+        Returns:
+            x (B, T, C), embed (B, T, C)  # embed is the token-only embedding
+        """
+        if token_embedding is None:
+            token_embedding = self.embed_tokens(src_tokens)
+        embed = token_embedding  # token-only
+
+        x = self.embed_scale * token_embedding
+        if self.embed_positions is not None:
+            x = x + self.embed_positions(src_tokens)
+        if self.layernorm_embedding is not None:
+            x = self.layernorm_embedding(x)
+        # No dropout, no quant_noise here (inference-only)
+        return x, embed
+
+    # ----------------------------------------------------------------------
+    def forward(
+        self,
+        src_tokens: Tensor,
+        src_lengths: Optional[Tensor] = None,
+        return_all_hiddens: bool = False,
+        token_embeddings: Optional[Tensor] = None,
+        *,
+        # External-inputs branch (used for export)
+        encoder_padding_mask: Optional[Tensor] = None,  # B x T (bool)
+    ) -> Tensor | Dict[str, List[Optional[Tensor]]]:
+        """
+        If `self.use_external_inputs` is True:
+          - Use only x_external and encoder_padding_mask.
+          - Return a single Tensor (T, B, C) for export friendliness.
+
+        Otherwise (False):
+          - Behave like the original Fairseq encoder forward and return dict-of-lists.
+        """
+        if self.use_external_inputs:
+            # ----- External-input mode: completely skip embedding/positional/LN/mask creation -----
+            x_external = src_tokens  # T x B x C (already embedded + transposed)
+
+            encoder_states: List[Tensor] = []
+            if return_all_hiddens:
+                encoder_states.append(x_external)
+
+            for layer in self.layers:
+                out = layer(x_external, encoder_padding_mask=encoder_padding_mask)
+                x_external = (
+                    out[0] if (isinstance(out, tuple) and len(out) == 2) else out
+                )
+                if return_all_hiddens:
+                    encoder_states.append(x_external)
+
+            if self.layer_norm is not None:
+                x_external = self.layer_norm(x_external)
+
+            if self.return_type == "dict":
+                return {
+                    "encoder_out": [x_external],
+                    "encoder_padding_mask": [encoder_padding_mask],
+                    "encoder_states": encoder_states,  # type: ignore[dict-item]
+                }
+            else:
+                # For export, returning a single Tensor is simpler and more portable.
+                return x_external
+
+        # ----- Original path (training/eval compatibility) ------------------
+
+        # Compute padding mask [B, T] (bool). We keep the original "has_pads" logic.
+        encoder_padding_mask = src_tokens.eq(self.padding_idx)
+        has_pads: Tensor = (
+            torch.tensor(src_tokens.device.type == "xla") or encoder_padding_mask.any()
+        )
+        if torch.jit.is_scripting():
+            has_pads = torch.tensor(1) if has_pads else torch.tensor(0)
+
+        # Embedding path (B,T,C). No dropout/quant_noise.
+        x, encoder_embedding = self.forward_embedding(src_tokens, token_embeddings)
+
+        # Zero out padded timesteps prior to the stack (same as original)
+        x = x * (
+            1 - encoder_padding_mask.unsqueeze(-1).type_as(x) * has_pads.type_as(x)
+        )
+
+        # B x T x C -> T x B x C
+        x = x.transpose(0, 1)
+
+        encoder_states: List[Tensor] = []  # type: ignore[no-redef]
+        fc_results: List[Optional[Tensor]] = []
+
+        if return_all_hiddens:
+            encoder_states.append(x)
+
+        # Encoder layers (each item is PTQ-wrapped and uses static additive masks internally)
+        for layer in self.layers:
+            out = layer(
+                x, encoder_padding_mask=encoder_padding_mask if has_pads else None
+            )
+            if isinstance(out, tuple) and len(out) == 2:
+                x, fc_res = out
+            else:
+                x = out
+                fc_res = None
+
+            if return_all_hiddens and not torch.jit.is_scripting():
+                encoder_states.append(x)
+                fc_results.append(fc_res)
+
+        if self.layer_norm is not None:
+            x = self.layer_norm(x)
+
+        # src_lengths (B, 1) int32, identical to original
+        src_lengths_out = (
+            src_tokens.ne(self.padding_idx)
+            .sum(dim=1, dtype=torch.int32)
+            .reshape(-1, 1)
+            .contiguous()
+        )
+
+        return {
+            "encoder_out": [x],  # T x B x C
+            "encoder_padding_mask": [encoder_padding_mask],  # B x T
+            "encoder_embedding": [encoder_embedding],  # B x T x C
+            "encoder_states": encoder_states,  # type: ignore[dict-item]  # List[T x B x C]
+            "fc_results": fc_results,  # type: ignore[dict-item]  # List[T x B x C]
+            "src_tokens": [],
+            "src_lengths": [src_lengths_out],
+        }
+
+    def forward_torchscript(self, net_input: Dict[str, Tensor]):
+        """A TorchScript-compatible version of forward.
+
+        Encoders which use additional arguments may want to override
+        this method for TorchScript compatibility.
+        """
+        if "encoder_padding_mask" in net_input:
+            return self.forward(
+                src_tokens=net_input["src_tokens"],
+                src_lengths=net_input["src_lengths"],
+                encoder_padding_mask=net_input["encoder_padding_mask"],
+            )
+        else:
+            return self.forward(
+                src_tokens=net_input["src_tokens"],
+                src_lengths=net_input["src_lengths"],
+            )
+
+    # ----------------------------------------------------------------------
+    @torch.jit.export
+    def reorder_encoder_out(self, encoder_out: Dict[str, List[Tensor]], new_order):
+        """
+        Match original API: reorder the batched dimension (B) according to new_order.
+        """
+        reordered = dict()  # type: ignore[var-annotated]
+        if len(encoder_out["encoder_out"]) == 0:
+            new_encoder_out = []
+        else:
+            new_encoder_out = [encoder_out["encoder_out"][0].index_select(1, new_order)]
+        reordered["encoder_out"] = new_encoder_out
+        keys = [
+            "encoder_padding_mask",
+            "encoder_embedding",
+            "src_tokens",
+            "src_lengths",
+        ]
+        for k in keys:
+            if k not in encoder_out:
+                continue
+            if len(encoder_out[k]) == 0:
+                reordered[k] = []
+            else:
+                reordered[k] = [encoder_out[k][0].index_select(0, new_order)]
+
+        if "encoder_states" in encoder_out:
+            encoder_states = encoder_out["encoder_states"]
+            if len(encoder_states) > 0:
+                for idx, state in enumerate(encoder_states):
+                    encoder_states[idx] = state.index_select(1, new_order)
+            reordered["encoder_states"] = encoder_states
+
+        return reordered
+
+    @torch.jit.export
+    def _reorder_encoder_out(self, encoder_out: Dict[str, List[Tensor]], new_order):
+        """Dummy re-order for beamable enc-dec attention (API parity)."""
+        return encoder_out
+
+    def max_positions(self) -> int:
+        """Maximum input length supported by the encoder (same policy as the original)."""
+        if self.embed_positions is None:
+            return self.max_source_positions
+        return min(self.max_source_positions, self.embed_positions.max_positions)
+
+    def upgrade_state_dict_named(self, state_dict, name):
+        """
+        Forward-compat mapping for older checkpoints (mirror original behavior for LNs).
+        The actual remapping of per-layer norms is delegated to the wrapped layers.
+        """
+        for i, layer in enumerate(self.layers):
+            if hasattr(layer, "upgrade_state_dict_named"):
+                layer.upgrade_state_dict_named(state_dict, f"{name}.layers.{i}")
+
+        version_key = f"{name}.version"
+        v = state_dict.get(version_key, torch.Tensor([1]))
+        if float(v[0].item()) < 2:
+            self.layer_norm = None
+            state_dict[version_key] = torch.Tensor([1])
+        return state_dict
+
+    def _all_observers(self):
+        for m in self.layers:
+            if isinstance(m, QuantModuleBase):
+                yield from m._all_observers()

tico/quantization/wrapq/wrappers/fairseq/quant_encoder_layer.py

@@ -0,0 +1,163 @@
+# 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.quantization.config.ptq import PTQConfig
+from tico.quantization.wrapq.wrappers.fairseq.quant_mha import (
+    QuantFairseqMultiheadAttention,
+)
+from tico.quantization.wrapq.wrappers.ptq_wrapper import PTQWrapper
+from tico.quantization.wrapq.wrappers.quant_module_base import QuantModuleBase
+from tico.quantization.wrapq.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[PTQConfig] = 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()