Upload 8 files

upload gbst-base-ds6x-newblock-1144k

config.json

@@ -0,0 +1,82 @@
+{
+  "_name_or_path": "/home/jhshin/gbst-base-ds6x-newblock-972k-240517/",
+  "architectures": [
+    "GBSWT5ForConditionalGeneration"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_gbswt5.GBSWT5Config",
+    "AutoModel": "modeling_gbswt5.GBSWT5ForModel",
+    "AutoModelForSeq2SeqLM": "modeling_gbswt5.GBSWT5ForConditionalGeneration"
+  },
+  "d_ff": 3968,
+  "d_kv": 64,
+  "d_model": 1536,
+  "decoder_start_token_id": 0,
+  "dense_act_fn": "gelu_new",
+  "downsample_factor": 6,
+  "dropout_rate": 0.0,
+  "eos_token_id": 1,
+  "feed_forward_proj": "gated-gelu",
+  "gbst_batchnorm": false,
+  "gradient_checkpointing": false,
+  "initializer_factor": 0.05,
+  "is_encoder_decoder": true,
+  "is_gated_act": true,
+  "kv_heads": null,
+  "layer_norm_epsilon": 1e-06,
+  "max_subword_block_size": null,
+  "model_type": "gbswt5",
+  "num_decoder_layers": 6,
+  "num_heads": 12,
+  "num_layers": 18,
+  "output_past": true,
+  "pad_token_id": 0,
+  "relative_attention_max_distance": 128,
+  "relative_attention_num_buckets": 32,
+  "score_consensus_attn": true,
+  "subword_blocks": [
+    [
+      1,
+      0
+    ],
+    [
+      2,
+      0
+    ],
+    [
+      3,
+      0
+    ],
+    [
+      4,
+      0
+    ],
+    [
+      5,
+      0
+    ],
+    [
+      6,
+      0
+    ],
+    [
+      7,
+      0
+    ],
+    [
+      8,
+      0
+    ],
+    [
+      9,
+      0
+    ]
+  ],
+  "tie_word_embeddings": false,
+  "tokenizer_class": "ByT5Tokenizer",
+  "torch_dtype": "float32",
+  "transformers_version": "4.33.2",
+  "use_cache": true,
+  "vocab_size": 384,
+  "z_loss": 0.0001
+}

configuration_gbswt5.py

@@ -0,0 +1,139 @@
+"""
+   GBSWT5 model configuration.
+
+   Copyright (C) 2023~ ETRI LIRS. Jong-hun Shin.
+"""
+
+from typing import Mapping
+from transformers.configuration_utils import PretrainedConfig
+from transformers.onnx import OnnxSeq2SeqConfigWithPast
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+_BLOCKS = (
+    (1, 0), (2, 0), (3, 0), (4, 0),
+    (6, 0), (9, 0),
+    #(12, 0), (12, 3), (12, 6), (12, 9)
+)
+
+
+class GBSWT5Config(PretrainedConfig):
+    """ Based on models.t5. configuration_t5. T5Config in hf Transformers. """
+    model_type = "gbswt5"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {"hidden_size": "d_model",
+                     "num_attention_heads": "num_heads",
+                     "num_hidden_layers": "num_layers"}
+
+    def __init__(
+        self,
+        vocab_size=384,
+        d_model=512,
+        d_kv=64,
+        d_ff=2048,
+        num_layers=6,
+        num_decoder_layers=None,
+        num_heads=8,
+        relative_attention_num_buckets=32,
+        relative_attention_max_distance=128,
+        dropout_rate=0.1,
+        layer_norm_epsilon=1e-6,
+        initializer_factor=1.0,
+        feed_forward_proj="relu",
+        is_encoder_decoder=True,
+        use_cache=True,
+        pad_token_id=0,
+        eos_token_id=1,
+        max_subword_block_size=None,            # GBSWT-related options here from
+        subword_blocks=_BLOCKS,
+        downsample_factor=1,
+        score_consensus_attn=True,
+        z_loss=1e-4,
+        gbst_batchnorm=False,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.d_model = d_model
+        self.d_kv = d_kv
+        self.d_ff = d_ff
+        self.num_layers = num_layers
+        self.num_decoder_layers = (
+            num_decoder_layers if num_decoder_layers is not None else self.num_layers
+        )  # default = symmetry
+        self.num_heads = num_heads
+        self.relative_attention_num_buckets = relative_attention_num_buckets
+        self.relative_attention_max_distance = relative_attention_max_distance
+        self.dropout_rate = dropout_rate
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_factor = initializer_factor
+        self.feed_forward_proj = feed_forward_proj
+        self.use_cache = use_cache
+
+        act_info = self.feed_forward_proj.split("-")
+        self.dense_act_fn = act_info[-1]
+        self.is_gated_act = act_info[0] == "gated"
+
+        # GBSWT-related configurations
+        self.max_subword_block_size = max_subword_block_size
+        self.subword_blocks = subword_blocks
+        self.downsample_factor = downsample_factor
+        self.score_consensus_attn = score_consensus_attn
+        self.gbst_batchnorm = gbst_batchnorm
+
+        # z_loss for computational stability.
+        # see https://github.com/tensorflow/mesh/blob \
+        #         /fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L666
+        # (1) logits이 0으로 부터 너무 멀어지게 드리프팅 되지 않도록 하여, bf16에서 발생하는
+        # round-off error를 방지하기 위함. (2) 로짓이 normalized log-probabilities가 되도록 제고한다.
+        self.z_loss = z_loss
+
+        if self.subword_blocks is not None and isinstance(self.subword_blocks, list):
+            for idx, elem in enumerate(self.subword_blocks):
+                self.subword_blocks[idx] = tuple(elem)
+            self.subword_blocks = tuple(self.subword_blocks)
+
+        if len(act_info) > 1 and act_info[0] != "gated" or len(act_info) > 2:
+            raise ValueError(
+                f"`feed_forward_proj`: {feed_forward_proj} is not a valid activation function of the dense layer."
+                "Please make sure `feed_forward_proj` is of the format `gated-{ACT_FN}` or `{ACT_FN}`, e.g. "
+                "'gated-gelu' or 'relu'"
+            )
+
+        # for backwards compatibility
+        if feed_forward_proj == "gated-gelu":
+            self.dense_act_fn = "gelu_new"
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            eos_token_id=eos_token_id,
+            is_encoder_decoder=is_encoder_decoder,
+            **kwargs,
+        )
+
+
+class GBSWT5OnnxConfig(OnnxSeq2SeqConfigWithPast):
+    """ just copy of T5OnnxConfig. """
+    @property
+    def inputs(self) -> Mapping[str, Mapping[int, str]]:
+        common_inputs = {
+            "input_ids": {0: "batch", 1: "encoder_sequence"},
+            "attention_mask": {0: "batch", 1: "encoder_sequence"},
+        }
+        if self.use_past:
+            common_inputs["attention_mask"][1] = "past_encoder_sequence + sequence"
+            common_inputs["decoder_input_ids"] = {0: "batch"}
+            common_inputs["decoder_attention_mask"] = {0: "batch", 1: "past_decoder_sequence + sequence"}
+        else:
+            common_inputs["decoder_input_ids"] = {0: "batch", 1: "decoder_sequence"}
+            common_inputs["decoder_attention_mask"] = {0: "batch", 1: "decoder_sequence"}
+
+        if self.use_past:
+            self.fill_with_past_key_values_(common_inputs, direction="inputs")
+
+        return common_inputs
+
+    @property
+    def default_onnx_opset(self) -> int:
+        return 13
+

gbst.py

@@ -0,0 +1,254 @@
+"""
+    Gradient-based Subword Tokenization(GBST) Layer implementation.
+
+    based on lucidrains/charformer-pytorch implementation,
+    which distributed under MIT License.
+
+    original code location:
+    https://github.com/lucidrains/charformer-pytorch/charformer_pytorch.py
+
+    copyright (c) 2023~, ETRI LIRS. Jong-hun Shin.
+"""
+import math
+import functools
+import torch
+import torch.nn.functional as F
+
+from typing import Optional
+
+from torch import einsum, nn, Tensor
+from transformers.utils import logging
+from einops.layers.torch import Rearrange
+from einops import rearrange, repeat
+
+
+logger = logging.get_logger(__name__)
+
+# Block definition
+_BLOCKS = (
+    (1, 0), (2, 0), (3, 0), (4, 0),
+    (6, 0), (9, 0),
+    #(12, 0), (12, 3), (12, 6), (12, 9)
+)
+
+@torch.jit.script
+def pad_to_multiple(in_tensor:Tensor, multiple:int, seq_dim:int,
+                    dim:int, value:Optional[float]):
+    seqlen = in_tensor.shape[seq_dim]
+    padded_len = math.ceil(seqlen / multiple) * multiple
+    if seqlen == padded_len:
+        return in_tensor
+    pad_offset = (0,) * (-1 - dim) * 2
+    if len(pad_offset) == 0:
+        return F.pad(in_tensor, (0, padded_len - seqlen), value=value)
+    # unpack 2 dims
+    d1, d2 = pad_offset
+    return F.pad(in_tensor, (d1, d2, 0, padded_len - seqlen), value=value)
+
+
+
+
+class Depthwise1dConv(nn.Module):
+    def __init__(self, in_dim, out_dim, krnl_size, use_bn=False):
+        super().__init__()
+        self.use_bn = use_bn
+        self.convol = nn.Conv1d(in_dim, out_dim, krnl_size, groups=in_dim)
+        # EXPERIMENTAL: add BatchNorm Layer
+        if self.use_bn:
+            self.bn = nn.BatchNorm1d(out_dim, eps=1e-05,)
+        self.proj = nn.Conv1d(out_dim, out_dim, 1)
+
+    @torch.cuda.amp.autocast(enabled=False, dtype=torch.float32)
+    def forward(self, in_tensor):
+        in_tensor = self.convol(in_tensor)
+        if self.use_bn:
+            in_tensor = self.bn(in_tensor)
+        return self.proj(in_tensor)
+
+    def _init_weights(self, factor:float=0.05):
+        logger.debug(f"1dConv-Weight initialize called, before: {self.convol.weight.data}")
+        self.convol.weight.data.normal_(mean=0.0, std=factor * 1.0)
+        self.proj.weight.data.normal_(mean=0.0, std=factor * 1.0)
+        logger.debug(f"1dConv-Weight initialize called, after: {self.convol.weight.data}")
+
+
+class Padding(nn.Module):
+    def __init__(self, padding, value=0):
+        super().__init__()
+        self.padding = padding
+        self.value = value
+
+    def forward(self, in_tensor):
+        return F.pad(in_tensor, self.padding, value=self.value)
+
+
+class GBSWT(nn.Module):
+    """ Gradient-based Sub-Word Tokenizer implementation. """
+    def __init__(self, embed_tokens,
+                 max_block_size=None,
+                 blocks=_BLOCKS,
+                 downsample_factor=1,
+                 score_consensus_attn=True,
+                 use_bn=False,):
+        super().__init__()
+        num_tokens, dim = embed_tokens.weight.shape
+
+        assert (max_block_size is not None) ^ (blocks is not None), \
+                'max_block_size or blocks must be given.'
+        if blocks is None:
+            self.blocks = tuple(map(lambda elem: (elem, 0), range(1, max_block_size+1)))
+        else:
+            if not isinstance(blocks, tuple):
+                raise ValueError('blocks must be assigned as a tuple')
+            self.blocks = tuple(map(lambda elem: elem if isinstance(elem, tuple) else (elem, 0), blocks))
+            if not all([(offset < block_size) for block_size, offset in self.blocks]):
+                raise ValueError('Offset must be smaller than given block size.')
+            max_block_size = max(list(map(lambda x: x[0], self.blocks)))
+
+        assert downsample_factor <= max_block_size, \
+            'downsample factor must be less than the max_block_size.'
+
+        self.downsample_factor = downsample_factor
+        self.score_consensus_attn = score_consensus_attn
+        self.use_bn = use_bn
+        logger.debug(f"GBSWT Subword Block Combinations: {self.blocks}")
+        logger.debug(f"GBSWT Downsampling factor: {self.downsample_factor}, use BatchNorm: {self.use_bn}")
+
+        def lcm(*num):
+            return int(functools.reduce(lambda x, y: int((x * y) / math.gcd(x, y)), num, 1))
+
+        self.block_pad_multiple = lcm(*[block_size for block_size, _ in self.blocks])
+        #print(f"block_pad_multiple: {self.block_pad_multiple}")
+
+        # layer definition
+        self.embeds = embed_tokens
+        self.positional_convol = nn.Sequential(
+            Padding((0, 0, 0, max_block_size-1)),
+            Rearrange('b s d -> b d s'),
+            Depthwise1dConv(dim, dim, krnl_size=max_block_size, use_bn=self.use_bn,),
+            Rearrange('b d s -> b s d'))
+        self.cand_scoring = nn.Sequential(
+            nn.Linear(dim, 1),
+            Rearrange('... () -> ...'))
+
+    def _init_weights(self, factor:float=0.05):
+        self.positional_convol[2]._init_weights(factor)
+        #print(f"GBSTW weight initialization called: before: {self.cand_scoring[0].weight.data}")
+        self.cand_scoring[0].weight.data.normal_(mean=0.0, std=factor * 1.0)
+        #print(f"GBSTW weight initialization called: after: {self.cand_scoring[0].weight.data}")
+
+    def get_blocks(self):
+        """ return GBST candidate blocking list. """
+        return self.blocks
+
+    @torch.cuda.amp.autocast()
+    def forward(self, in_tensor, attention_mask=None):
+        b, s = in_tensor.shape
+        #print(f"initial shape: b, s : {b}, {s}, in_tensor.shape: {in_tensor.shape}")
+        mask = attention_mask
+        #print(f"mask: {mask}")
+        block_multi, ds_factor = self.block_pad_multiple, self.downsample_factor
+
+        in_tensor = self.embeds(in_tensor)
+        in_tensor = self.positional_convol(in_tensor)
+        in_tensor = pad_to_multiple(in_tensor, block_multi,
+                                    seq_dim=1, dim=-2, value=0.0)
+        if mask is not None:
+            mask = pad_to_multiple(mask, block_multi,
+                                   seq_dim=1, dim=-1, value=False)
+
+        def _masked_mean(in_tensor:Tensor, mask:Tensor, dim:int=-1):
+            len_diff = len(in_tensor.shape) - len(mask.shape)
+            mask = torch.unsqueeze(mask, dim=-len_diff)
+            in_tensor.masked_fill_(~(mask.bool()), 0.)
+
+            total_elems = mask.sum(dim=dim)
+            mean = in_tensor.sum(dim=dim) / total_elems.clamp(min=1.)
+            mean.masked_fill_((total_elems == 0), 0.)
+            return mean.float()
+
+        block_reprs, block_masks = [], []
+
+        # 이제 입력 시퀀스를 cloning해서 후보를 세팅
+        for block_size, offset in self.blocks:
+            block_in = in_tensor.clone()
+            if mask is not None:
+                block_mask = mask.clone()
+            need_padding = offset > 0
+
+            if need_padding:
+                loff, roff = (block_size - offset), offset
+                #print(f"loff: {loff}, roff: {roff}")
+                block_in = F.pad(block_in, (0, 0, loff, roff), value=0.0)
+                if mask is not None:
+                    block_mask = F.pad(block_mask, (0, 0, loff, roff), value=False)
+
+            blks = rearrange(block_in, 'b (s m) d -> b s m d', m=block_size)
+            if mask is not None:
+                mask_blks = rearrange(block_mask, 'b (s m) -> b s m', m=block_size)
+                blk_repr = _masked_mean(blks, mask_blks, dim=-2)
+            else:
+                blk_repr = blks.mean(dim=-2)
+
+            blk_repr = repeat(blk_repr, 'b s d -> b (s m) d', m=block_size)
+
+            if need_padding:
+                blk_repr = blk_repr[:, loff:-roff]
+
+            block_reprs.append(blk_repr)
+
+            if mask is not None:
+                mask_blks = torch.any(mask_blks, dim=-1)
+                mask_blks = repeat(mask_blks, 'b s -> b (s m)', m=block_size)
+                if need_padding:
+                    mask_blks = mask_blks[:, loff:-roff]
+                block_masks.append(mask_blks)
+
+        # stack them all
+        block_reprs = torch.stack(block_reprs, dim=2,)
+        scores = self.cand_scoring(block_reprs)
+
+        if mask is not None:
+            block_masks = torch.stack(block_masks, dim=2)
+            max_neg_val = -torch.finfo(scores.dtype).max
+            scores = scores.masked_fill(~block_masks, max_neg_val)
+
+        scores = scores.softmax(dim=2)
+
+        # cheap consensus attention, as equation (5) in paper.
+        if self.score_consensus_attn:
+            score_sim = einsum('b i d, b j d -> b i j', scores, scores)
+
+            if mask is not None:
+                cross_mask = rearrange(mask, 'b i -> b i ()') * rearrange(mask, 'b j -> b () j')
+                max_neg_val = -torch.finfo(score_sim.dtype).max
+                score_sim = score_sim.masked_fill((~(cross_mask.bool())), max_neg_val)
+
+            score_attn = score_sim.softmax(dim=-1)
+            scores = einsum('b i j, b j m -> b i m', score_attn, scores)
+
+        scores = rearrange(scores, 'b n m -> b n m ()')
+        in_tensor = (block_reprs * scores).sum(dim=2)
+
+        @torch.jit.script
+        def _reshape_input_tensor(in_tensor:Tensor, s:int, d:int):
+            # get divisible length to pad
+            m = int(math.ceil(s / d) * d)
+            #print(f"_reshape_input_tensor: {m}")
+            return in_tensor[:, :m]
+
+        in_tensor = _reshape_input_tensor(in_tensor, s, ds_factor)
+        if mask is not None:
+            mask = _reshape_input_tensor(mask, s, ds_factor)
+
+        # downsample with mean pooling
+        in_tensor = rearrange(in_tensor, 'b (n m) d -> b n m d', m=ds_factor)
+        if mask is not None:
+            mask = rearrange(mask, 'b (n m) -> b n m', m=ds_factor)
+            in_tensor = _masked_mean(in_tensor, mask, dim=2)
+            mask = torch.any(mask, dim=-1)
+        else:
+            in_tensor = in_tensor.mean(dim=-2)
+
+        # tuple을 반환하기 때문에, forward()에서 [0]을 취해 바꿔줘야 한다
+        return in_tensor, mask

generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "decoder_start_token_id": 0,
+  "eos_token_id": 1,
+  "pad_token_id": 0,
+  "transformers_version": "4.33.2"
+}

modeling_gbswt5.py

@@ -0,0 +1,766 @@
+"""
+    hf transformers-compatible GBST + T5 Model implementation.
+
+    several methods are copying from huggingface/transformers/models/t5/modeling_t5.py
+    as Implementation Standards for compatibility. (version 4.28.1)
+
+    hf transformers' modeling_t5.py file is distributed under Apache 2.0 License.
+
+    Copyright (C) 2023, ETRI LIRS, Jong-hun Shin.
+"""
+import copy
+
+from typing import Optional, Union, Tuple
+
+import torch
+
+from torch import nn
+from transformers import add_start_docstrings
+from transformers.utils import logging
+from transformers.modeling_outputs import (
+    BaseModelOutput,
+    BaseModelOutputWithPastAndCrossAttentions,
+    Seq2SeqLMOutput,
+    Seq2SeqModelOutput,
+)
+from transformers.models.t5.modeling_t5 import (
+    T5LayerNorm, T5Block, T5Stack,
+    T5Model, T5PreTrainedModel, T5ForConditionalGeneration, T5EncoderModel,
+    T5DenseActDense, T5DenseGatedActDense, T5Attention,
+    T5_START_DOCSTRING
+)
+
+from .configuration_gbswt5 import GBSWT5Config
+from .gbst import GBSWT
+
+
+logger = logging.get_logger(__name__)
+
+
+class GBSWT5PreTrainedModel(T5PreTrainedModel):
+    config_class = GBSWT5Config
+    base_model_prefix = "GBSWT5"
+    is_parallelizable = True
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["T5Block"]
+    _keep_in_fp32_modules = ["wo"]
+
+    def _init_weights(self, module):
+        """Initialize the weights. 대부분은 T5PreTrainedModel을 따른다. """
+        factor = self.config.initializer_factor  # Used for testing weights initialization
+        if isinstance(module, T5LayerNorm):
+            module.weight.data.fill_(factor * 1.0)
+        elif isinstance(
+            module,
+            ( GBSWT5Model, GBSWT5ForConditionalGeneration, GBSWT5EncoderModel,),
+        ):
+            # Mesh TensorFlow embeddings initialization
+            # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L1624
+            module.shared.weight.data.normal_(mean=0.0, std=factor * 1.0)
+            if hasattr(module, "lm_head") and not self.config.tie_word_embeddings:
+                module.lm_head.weight.data.normal_(mean=0.0, std=factor * 1.0)
+            if hasattr(module, "qa_outputs"):
+                module.qa_outputs.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
+                module.qa_outputs.bias.data.zero_()
+        elif isinstance(module, T5DenseActDense):
+            # Mesh TensorFlow FF initialization
+            # See https://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/transformer/transformer_layers.py#L56
+            # and https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L89
+            module.wi.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
+            if hasattr(module.wi, "bias") and module.wi.bias is not None:
+                module.wi.bias.data.zero_()
+            module.wo.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_ff) ** -0.5))
+            if hasattr(module.wo, "bias") and module.wo.bias is not None:
+                module.wo.bias.data.zero_()
+        elif isinstance(module, T5DenseGatedActDense):
+            module.wi_0.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
+            if hasattr(module.wi_0, "bias") and module.wi_0.bias is not None:
+                module.wi_0.bias.data.zero_()
+            module.wi_1.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
+            if hasattr(module.wi_1, "bias") and module.wi_1.bias is not None:
+                module.wi_1.bias.data.zero_()
+            module.wo.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_ff) ** -0.5))
+            if hasattr(module.wo, "bias") and module.wo.bias is not None:
+                module.wo.bias.data.zero_()
+        elif isinstance(module, T5Attention):
+            # Mesh TensorFlow attention initialization to avoid scaling before softmax
+            # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/attention.py#L136
+            d_model = self.config.d_model
+            key_value_proj_dim = self.config.d_kv
+            n_heads = self.config.num_heads
+            module.q.weight.data.normal_(mean=0.0, std=factor * ((d_model * key_value_proj_dim) ** -0.5))
+            module.k.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5))
+            module.v.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5))
+            module.o.weight.data.normal_(mean=0.0, std=factor * ((n_heads * key_value_proj_dim) ** -0.5))
+            if module.has_relative_attention_bias:
+                module.relative_attention_bias.weight.data.normal_(mean=0.0, std=factor * ((d_model) ** -0.5))
+        elif isinstance(module, GBSWT):
+            module._init_weights(factor)
+
+
+class GBSWT5Stack(GBSWT5PreTrainedModel):
+    """ implement GBST-enabled T5Model, based on HF Transformers's T5Stack. """
+    def __init__(self, config: GBSWT5Config, embed_tokens :nn.Embedding=None):
+        # 초기화는 이전의 것을 따른다. 상속이 좀 애매해서, 사실 별도로 정의해야 하나 싶기도 하다.
+        super().__init__(config)
+
+        # override embed_tokens, apply GBWST
+        self.embed_tokens = GBSWT(embed_tokens=embed_tokens,
+                                  max_block_size=config.max_subword_block_size,
+                                  blocks=config.subword_blocks,
+                                  downsample_factor=config.downsample_factor,
+                                  score_consensus_attn=config.score_consensus_attn,
+                                  use_bn=config.gbst_batchnorm,)
+        self.is_decoder = config.is_decoder
+
+        self.block = nn.ModuleList(
+            [T5Block(config, has_relative_attention_bias=bool(i == 0)) for i in range(config.num_layers)]
+        )
+        self.final_layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+        self.dropout = nn.Dropout(config.dropout_rate)
+
+        # Initialize weights and apply final processing, same as T5 Stack.
+        self.post_init()
+        # for Model Parallel
+        self.model_parallel = False
+        self.device_map = False
+        self.gradient_checkpointing = False
+        self.downsample_factor = config.downsample_factor
+
+    def forward(self,
+                input_ids=None,
+                attention_mask=None,
+                encoder_hidden_states=None,
+                encoder_attention_mask=None,
+                inputs_embeds=None,
+                head_mask=None,
+                cross_attn_head_mask=None,
+                past_key_values=None,
+                use_cache=None,
+                output_attentions=None,
+                output_hidden_states=None,
+                return_dict=None,
+                ):
+        """ GBST 파트를 제외하면, T5Stack.forward() 구현을 그대로 복제하였다. """
+        # Model parallel
+        if self.model_parallel:
+            torch.cuda.set_device(self.first_device)
+            self.embed_tokens = self.embed_tokens.to(self.first_device)
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is not None and inputs_embeds is not None:
+            err_msg_prefix = "decoder_" if self.is_decoder else ""
+            raise ValueError(
+                f"You cannot specify both {err_msg_prefix}input_ids and {err_msg_prefix}inputs_embeds at the same time"
+            )
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            err_msg_prefix = "decoder_" if self.is_decoder else ""
+            raise ValueError(f"You have to specify either {err_msg_prefix}input_ids or {err_msg_prefix}inputs_embeds")
+
+        if inputs_embeds is None:
+            assert self.embed_tokens is not None, "You have to initialize the model with valid token embeddings"
+            #print(f"old: {input_shape}")
+            inputs_embeds, attention_mask = self.embed_tokens(input_ids, attention_mask)
+            # for downsample_factor > 1
+            input_shape = inputs_embeds.size()[:-1]
+            #print(f"new: {input_shape}")
+
+        batch_size, seq_length = input_shape
+        #print(f"bs: {batch_size}, sl: {seq_length}")
+
+        # required mask seq length can be calculated via length of past
+        mask_seq_length = past_key_values[0][0].shape[2] + seq_length if past_key_values is not None else seq_length
+        #print(f"mask_seq_length: {mask_seq_length}")
+
+        if use_cache is True:
+            assert self.is_decoder, f"`use_cache` can only be set to `True` if {self} is used as a decoder"
+
+        if attention_mask is None:
+            attention_mask = torch.ones(batch_size, mask_seq_length, device=inputs_embeds.device)
+        if self.is_decoder and encoder_attention_mask is None and encoder_hidden_states is not None:
+            encoder_seq_length = encoder_hidden_states.shape[1]
+            encoder_attention_mask = torch.ones(
+                batch_size, encoder_seq_length, device=inputs_embeds.device, dtype=torch.long
+            )
+
+        # initialize past_key_values with `None` if past does not exist
+        if past_key_values is None:
+            past_key_values = [None] * len(self.block)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask = self.get_extended_attention_mask(attention_mask, input_shape)
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if self.is_decoder and encoder_hidden_states is not None:
+            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=inputs_embeds.device)
+            encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+        else:
+            encoder_extended_attention_mask = None
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        # Prepare head mask if needed
+        head_mask = self.get_head_mask(head_mask, self.config.num_layers)
+        cross_attn_head_mask = self.get_head_mask(cross_attn_head_mask, self.config.num_layers)
+        present_key_value_states = () if use_cache else None
+        all_hidden_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+        all_cross_attentions = () if (output_attentions and self.is_decoder) else None
+        position_bias = None
+        encoder_decoder_position_bias = None
+
+        hidden_states = self.dropout(inputs_embeds)
+
+        for i, (layer_module, past_key_value) in enumerate(zip(self.block, past_key_values)):
+            layer_head_mask = head_mask[i]
+            cross_attn_layer_head_mask = cross_attn_head_mask[i]
+            # Model parallel
+            if self.model_parallel:
+                torch.cuda.set_device(hidden_states.device)
+                # Ensure that attention_mask is always on the same device as hidden_states
+                if attention_mask is not None:
+                    attention_mask = attention_mask.to(hidden_states.device)
+                if position_bias is not None:
+                    position_bias = position_bias.to(hidden_states.device)
+                if encoder_hidden_states is not None:
+                    encoder_hidden_states = encoder_hidden_states.to(hidden_states.device)
+                if encoder_extended_attention_mask is not None:
+                    encoder_extended_attention_mask = encoder_extended_attention_mask.to(hidden_states.device)
+                if encoder_decoder_position_bias is not None:
+                    encoder_decoder_position_bias = encoder_decoder_position_bias.to(hidden_states.device)
+                if layer_head_mask is not None:
+                    layer_head_mask = layer_head_mask.to(hidden_states.device)
+                if cross_attn_layer_head_mask is not None:
+                    cross_attn_layer_head_mask = cross_attn_layer_head_mask.to(hidden_states.device)
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return tuple(module(*inputs, use_cache, output_attentions))
+
+                    return custom_forward
+
+                layer_outputs = checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    extended_attention_mask,
+                    position_bias,
+                    encoder_hidden_states,
+                    encoder_extended_attention_mask,
+                    encoder_decoder_position_bias,
+                    layer_head_mask,
+                    cross_attn_layer_head_mask,
+                    None,  # past_key_value is always None with gradient checkpointing
+                )
+            else:
+                layer_outputs = layer_module(
+                    hidden_states,
+                    attention_mask=extended_attention_mask,
+                    position_bias=position_bias,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_extended_attention_mask,
+                    encoder_decoder_position_bias=encoder_decoder_position_bias,
+                    layer_head_mask=layer_head_mask,
+                    cross_attn_layer_head_mask=cross_attn_layer_head_mask,
+                    past_key_value=past_key_value,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                )
+
+            # layer_outputs is a tuple with:
+            # hidden-states, key-value-states, (self-attention position bias), (self-attention weights), (cross-attention position bias), (cross-attention weights)
+            if use_cache is False:
+                layer_outputs = layer_outputs[:1] + (None,) + layer_outputs[1:]
+
+            hidden_states, present_key_value_state = layer_outputs[:2]
+
+            # We share the position biases between the layers - the first layer store them
+            # layer_outputs = hidden-states, key-value-states (self-attention position bias), (self-attention weights),
+            # (cross-attention position bias), (cross-attention weights)
+            position_bias = layer_outputs[2]
+            if self.is_decoder and encoder_hidden_states is not None:
+                encoder_decoder_position_bias = layer_outputs[4 if output_attentions else 3]
+            # append next layer key value states
+            if use_cache:
+                present_key_value_states = present_key_value_states + (present_key_value_state,)
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[3],)
+                if self.is_decoder:
+                    all_cross_attentions = all_cross_attentions + (layer_outputs[5],)
+
+            # Model Parallel: If it's the last layer for that device, put things on the next device
+            if self.model_parallel:
+                for k, v in self.device_map.items():
+                    if i == v[-1] and "cuda:" + str(k) != self.last_device:
+                        hidden_states = hidden_states.to("cuda:" + str(k + 1))
+
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+
+        # Add last layer
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    present_key_value_states,
+                    all_hidden_states,
+                    all_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            ), attention_mask
+
+        # must be return downsampled attention_mask
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=present_key_value_states,
+            hidden_states=all_hidden_states,
+            attentions=all_attentions,
+            cross_attentions=all_cross_attentions,
+        ), attention_mask
+
+    def get_input_embeddings(self):
+        return self.embed_tokens.embeds
+
+    def set_input_embeddings(self, new_embeddings):
+        self.embed_tokens.embeds = new_embeddings
+
+
+GBSWT5Stack.parallelize = T5Stack.parallelize
+GBSWT5Stack.deparallelize = T5Stack.deparallelize
+
+
+class GBSWT5Model(GBSWT5PreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [
+        "decoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weight",
+    ]
+    _tied_weights_keys = ["encoder.embed_tokens.embeds.weight", "decoder_embed_tokens.embeds.weight"]
+
+    def __init__(self, config: GBSWT5Config):
+        """ override T5Model """
+        # override some default missing parameters for pretrained ByT5 models (e.g. google/byt5-small)
+        if not hasattr(config, 'max_subword_block_size'):
+            config.max_subword_block_size = None
+        if not hasattr(config, 'subword_blocks'):
+            config.subword_blocks = ((1, 0), (2, 0), (3, 0), (6, 0), (9, 0),)
+        if not hasattr(config, 'downsample_factor'):
+            config.downsample_factor = 1
+        if not hasattr(config, 'score_consensus_attn'):
+            config.score_consensus_attn = True
+
+        super().__init__(config)
+
+        # naive T5와 같이 embedding은 공유함
+        self.shared = nn.Embedding(config.vocab_size, config.d_model)
+
+        encoder_cfg = copy.deepcopy(config)
+        encoder_cfg.is_decoder = False
+        encoder_cfg.use_cache = False
+        encoder_cfg.is_encoder_decoder = False
+        self.encoder = GBSWT5Stack(encoder_cfg, self.shared)
+
+        # Embedding base를 공유하기는 하지만, decoder에는 GBSWT를
+        # 적용하지 않아야 한다.
+        decoder_cfg = copy.deepcopy(config)
+        decoder_cfg.is_decoder = True
+        decoder_cfg.is_encoder_decoder = False
+        decoder_cfg.num_layers = config.num_decoder_layers
+        self.decoder = T5Stack(decoder_cfg, self.shared)
+
+        self.post_init()
+
+        self.model_parallel = False
+        self.device_map = None
+
+    def forward(self,
+                input_ids: Optional[torch.LongTensor] = None,
+                attention_mask: Optional[torch.FloatTensor] = None,
+                decoder_input_ids: Optional[torch.LongTensor] = None,
+                decoder_attention_mask: Optional[torch.BoolTensor] = None,
+                head_mask: Optional[torch.FloatTensor] = None,
+                decoder_head_mask: Optional[torch.FloatTensor] = None,
+                cross_attn_head_mask: Optional[torch.Tensor] = None,
+                encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+                past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+                inputs_embeds: Optional[torch.Tensor] = None,
+                decoder_inputs_embeds: Optional[torch.Tensor] = None,
+                use_cache: Optional[bool] = None,
+                output_attentions: Optional[bool] = None,
+                output_hidden_states: Optional[bool] = None,
+                return_dict: Optional[bool] = None,
+                ) -> Union[Tuple[torch.FloatTensor], Seq2SeqModelOutput]:
+        """
+        중요한 것은, downsampling이 된 경우 attention_mask가 변경되므로,
+        이를 반영해주는 것이 필요하다. hf transformers 4.29.1에서 복제함
+        """
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # FutureWarning: head_mask was separated into two input args - head_mask, decoder_head_mask
+        if head_mask is not None and decoder_head_mask is None:
+            if self.config.num_layers == self.config.num_decoder_layers:
+                warnings.warn(__HEAD_MASK_WARNING_MSG, FutureWarning)
+                decoder_head_mask = head_mask
+
+        # Encode if needed (training, first prediction pass)
+        if encoder_outputs is None:
+            encoder_outputs, attention_mask = self.encoder(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                inputs_embeds=inputs_embeds,
+                head_mask=head_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
+            # inference mode (e.g. .generate()) - must dewrap encoder output 'tuple'
+            encoder_outputs, attention_mask = encoder_outputs
+            encoder_outputs = BaseModelOutput(
+                last_hidden_state=encoder_outputs[0],
+                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
+                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
+            )
+
+        hidden_states = encoder_outputs[0]
+
+        # Set device for model parallelism
+        if self.model_parallel:
+            torch.cuda.set_device(self.decoder.first_device)
+            hidden_states = hidden_states.to(self.decoder.first_device)
+            if decoder_input_ids is not None:
+                decoder_input_ids = decoder_input_ids.to(self.decoder.first_device)
+            if attention_mask is not None:
+                attention_mask = attention_mask.to(self.decoder.first_device)
+            if decoder_attention_mask is not None:
+                decoder_attention_mask = decoder_attention_mask.to(self.decoder.first_device)
+
+        # Decode
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            inputs_embeds=decoder_inputs_embeds,
+            past_key_values=past_key_values,
+            encoder_hidden_states=hidden_states,
+            encoder_attention_mask=attention_mask,
+            head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        if not return_dict:
+            return decoder_outputs + encoder_outputs
+
+        return Seq2SeqModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            past_key_values=decoder_outputs.past_key_values,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+
+GBSWT5Model.parallelize = T5Model.parallelize
+GBSWT5Model.deparallelize = T5Model.deparallelize
+GBSWT5Model.get_input_embeddings = T5Model.get_input_embeddings
+GBSWT5Model.set_input_embeddings = T5Model.set_input_embeddings
+GBSWT5Model.get_encoder = T5Model.get_encoder
+GBSWT5Model._prune_heads = T5Model._prune_heads
+
+
+@add_start_docstrings("""T5 Model with a `language modeling` head on top.""", T5_START_DOCSTRING)
+class GBSWT5ForConditionalGeneration(GBSWT5PreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [
+        "decoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weight",
+    ]
+    _tied_weights_keys = ["encoder.embed_tokens.embeds.weight",
+                          "decoder_embed_tokens.embeds.weight",
+                          "lm_head.weight"]
+
+    def __init__(self, config: GBSWT5Config):
+        # override some default missing parameters for pretrained ByT5 models (e.g. google/byt5-small)
+        if not hasattr(config, 'max_subword_block_size'):
+            config.max_subword_block_size = None
+        if not hasattr(config, 'subword_blocks'):
+            config.subword_blocks = ((1, 0), (2, 0), (3, 0), (6, 0), (9, 0),)
+        if not hasattr(config, 'downsample_factor'):
+            config.downsample_factor = 1
+        if not hasattr(config, 'score_consensus_attn'):
+            config.score_consensus_attn = True
+
+        # Grandparent의 init를 그대로 상속, 나머지는 T5ForConditionalGeneration을 따름
+        super().__init__(config)
+
+        self.model_dim = config.d_model
+
+        self.shared = nn.Embedding(config.vocab_size, config.d_model)
+
+        encoder_cfg = copy.deepcopy(config)
+        encoder_cfg.is_decoder = False
+        encoder_cfg.use_cache = False
+        encoder_cfg.is_encoder_decoder = False
+        self.encoder = GBSWT5Stack(encoder_cfg, self.shared)
+
+        # Embedding base를 공유하기는 하지만, decoder에는 GBSWT를
+        # 적용하지 않아야 한다.
+        decoder_cfg = copy.deepcopy(config)
+        decoder_cfg.is_decoder = True
+        decoder_cfg.is_encoder_decoder = False
+        decoder_cfg.num_layers = config.num_decoder_layers
+        self.decoder = T5Stack(decoder_cfg, self.shared)
+
+        self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+    def forward(self,
+                input_ids: Optional[torch.LongTensor] = None,
+                attention_mask: Optional[torch.FloatTensor] = None,
+                decoder_input_ids: Optional[torch.LongTensor] = None,
+                decoder_attention_mask: Optional[torch.BoolTensor] = None,
+                head_mask: Optional[torch.FloatTensor] = None,
+                decoder_head_mask: Optional[torch.FloatTensor] = None,
+                cross_attn_head_mask: Optional[torch.Tensor] = None,
+                encoder_outputs: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+                past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+                inputs_embeds: Optional[torch.FloatTensor] = None,
+                decoder_inputs_embeds: Optional[torch.FloatTensor] = None,
+                labels: Optional[torch.LongTensor] = None,
+                use_cache: Optional[bool] = None,
+                output_attentions: Optional[bool] = None,
+                output_hidden_states: Optional[bool] = None,
+                return_dict: Optional[bool] = None,
+                ) -> Union[Tuple[torch.FloatTensor], Seq2SeqLMOutput]:
+        """
+        중요한 것은 encoder outputs에서 수정된 attention_mask를 다시 반영해야 하는 것임
+        downsampling이 들어간 경우, attention_mask가 변경되기 때문.
+        """
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # FutureWarning: head_mask was separated into two input args - head_mask, decoder_head_mask
+        if head_mask is not None and decoder_head_mask is None:
+            if self.config.num_layers == self.config.num_decoder_layers:
+                warnings.warn(__HEAD_MASK_WARNING_MSG, FutureWarning)
+                decoder_head_mask = head_mask
+
+        # Encode if needed (training, first prediction pass)
+        if encoder_outputs is None:
+            # Convert encoder inputs in embeddings if needed
+            encoder_outputs, attention_mask = self.encoder(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                inputs_embeds=inputs_embeds,
+                head_mask=head_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
+            # inference mode (e.g. .generate()) - must dewrap encoder output 'tuple'
+            encoder_outputs, attention_mask = encoder_outputs
+            encoder_outputs = BaseModelOutput(
+                last_hidden_state=encoder_outputs[0],
+                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
+                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
+            )
+
+        hidden_states = encoder_outputs[0]
+
+        if self.model_parallel:
+            torch.cuda.set_device(self.decoder.first_device)
+
+        if labels is not None and decoder_input_ids is None and decoder_inputs_embeds is None:
+            # get decoder inputs from shifting lm labels to the right
+            decoder_input_ids = self._shift_right(labels)
+
+        # Set device for model parallelism
+        if self.model_parallel:
+            torch.cuda.set_device(self.decoder.first_device)
+            hidden_states = hidden_states.to(self.decoder.first_device)
+            if decoder_input_ids is not None:
+                decoder_input_ids = decoder_input_ids.to(self.decoder.first_device)
+            if attention_mask is not None:
+                attention_mask = attention_mask.to(self.decoder.first_device)
+            if decoder_attention_mask is not None:
+                decoder_attention_mask = decoder_attention_mask.to(self.decoder.first_device)
+
+        # Decode
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            inputs_embeds=decoder_inputs_embeds,
+            past_key_values=past_key_values,
+            encoder_hidden_states=hidden_states,
+            encoder_attention_mask=attention_mask,
+            head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = decoder_outputs[0]
+
+        # Set device for model parallelism
+        if self.model_parallel:
+            torch.cuda.set_device(self.encoder.first_device)
+            self.lm_head = self.lm_head.to(self.encoder.first_device)
+            sequence_output = sequence_output.to(self.lm_head.weight.device)
+
+        if self.config.tie_word_embeddings:
+            # Rescale output before projecting on vocab
+            # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/transformer.py#L586
+            sequence_output = sequence_output * (self.model_dim**-0.5)
+
+        lm_logits = self.lm_head(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
+            # move labels to correct device to enable PP
+            labels = labels.to(lm_logits.device)
+            loss = loss_fct(lm_logits.view(-1, lm_logits.size(-1)), labels.view(-1))
+            # add z_loss for computational stability in bf16 amp.
+            # see https://github.com/huggingface/transformers/pull/10956#issuecomment-820712267
+            if self.config.z_loss != 0.0:
+                log_z = lm_logits.view(-1).logsumexp(-1)
+                loss += self.config.z_loss * log_z.square()
+
+        if not return_dict:
+            output = (lm_logits,) + decoder_outputs[1:] + encoder_outputs
+            return ((loss,) + output) if loss is not None else output
+
+        return Seq2SeqLMOutput(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=decoder_outputs.past_key_values,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+
+GBSWT5ForConditionalGeneration.parallelize = T5ForConditionalGeneration.parallelize
+GBSWT5ForConditionalGeneration.deparallelize = T5ForConditionalGeneration.deparallelize
+GBSWT5ForConditionalGeneration.get_input_embeddings = T5ForConditionalGeneration.get_input_embeddings
+GBSWT5ForConditionalGeneration.set_input_embeddings = T5ForConditionalGeneration.set_input_embeddings
+GBSWT5ForConditionalGeneration.get_output_embeddings = T5ForConditionalGeneration.get_output_embeddings
+GBSWT5ForConditionalGeneration.set_output_embeddings = T5ForConditionalGeneration.set_output_embeddings
+GBSWT5ForConditionalGeneration.get_encoder = T5ForConditionalGeneration.get_encoder
+GBSWT5ForConditionalGeneration.prepare_inputs_for_generation = T5ForConditionalGeneration.prepare_inputs_for_generation
+GBSWT5ForConditionalGeneration.prepare_decoder_input_ids_from_labels = T5ForConditionalGeneration.prepare_decoder_input_ids_from_labels
+GBSWT5ForConditionalGeneration._reorder_cache = T5ForConditionalGeneration._reorder_cache
+GBSWT5ForConditionalGeneration._prune_heads = T5Model._prune_heads
+
+
+class GBSWT5EncoderModel(T5PreTrainedModel):
+    _tied_weights_keys = ["encoder.embed_tokens.embeds.weight"]
+
+    def __init__(self, config: GBSWT5Config):
+        # override some default missing parameters for pretrained ByT5 models (e.g. google/byt5-small)
+        if not hasattr(config, 'max_subword_block_size'):
+            config.max_subword_block_size = None
+        if not hasattr(config, 'subword_blocks'):
+            config.subword_blocks = ((1, 0), (2, 0), (3, 0), (6, 0), (9, 0),)
+        if not hasattr(config, 'downsample_factor'):
+            config.downsample_factor = 1
+        if not hasattr(config, 'score_consensus_attn'):
+            config.score_consensus_attn = True
+
+        super().__init__(config)
+
+        # naive T5와 같이 embedding은 공유함
+        self.shared = nn.Embedding(config.vocab_size, config.d_model)
+
+        encoder_cfg = copy.deepcopy(config)
+        encoder_cfg.is_decoder = False
+        encoder_cfg.use_cache = False
+        encoder_cfg.is_encoder_decoder = False
+        self.encoder = GBSWT5Stack(encoder_cfg, self.shared)
+
+        self.post_init()
+
+        self.model_parallel = False
+        self.device_map = None
+
+    def forward(self,
+                input_ids: Optional[torch.LongTensor] = None,
+                attention_mask: Optional[torch.FloatTensor] = None,
+                head_mask: Optional[torch.FloatTensor] = None,
+                inputs_embeds: Optional[torch.FloatTensor] = None,
+                output_attentions: Optional[bool] = None,
+                output_hidden_states: Optional[bool] = None,
+                return_dict: Optional[bool] = None,
+                return_resized_attention_mask: Optional[bool] = None,
+                ) -> Union[Tuple[torch.FloatTensor], BaseModelOutput]:
+        r"""
+        downsampled 된 attention_mask를 함께 반환한다. 단, return_resized_attention_mask=True일 때만.
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        encoder_outputs, attention_mask = self.encoder(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        if return_resized_attention_mask:
+            return encoder_outputs, attention_mask
+
+        return encoder_outputs
+
+
+GBSWT5EncoderModel.parallelize = T5EncoderModel.parallelize
+GBSWT5EncoderModel.deparallelize = T5EncoderModel.deparallelize
+GBSWT5EncoderModel.get_input_embeddings = T5EncoderModel.get_input_embeddings
+GBSWT5EncoderModel.set_input_embeddings = T5EncoderModel.set_input_embeddings
+GBSWT5EncoderModel.get_encoder = T5EncoderModel.get_encoder
+GBSWT5EncoderModel._prune_heads = T5EncoderModel._prune_heads

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c90684b6b4905ad58b09a39183a3651f2c969f1c584cf5733886c2239d7519a4
+size 2336209429

special_tokens_map.json

@@ -0,0 +1,150 @@
+{
+  "additional_special_tokens": [
+    "<extra_id_0>",
+    "<extra_id_1>",
+    "<extra_id_2>",
+    "<extra_id_3>",
+    "<extra_id_4>",
+    "<extra_id_5>",
+    "<extra_id_6>",
+    "<extra_id_7>",
+    "<extra_id_8>",
+    "<extra_id_9>",
+    "<extra_id_10>",
+    "<extra_id_11>",
+    "<extra_id_12>",
+    "<extra_id_13>",
+    "<extra_id_14>",
+    "<extra_id_15>",
+    "<extra_id_16>",
+    "<extra_id_17>",
+    "<extra_id_18>",
+    "<extra_id_19>",
+    "<extra_id_20>",
+    "<extra_id_21>",
+    "<extra_id_22>",
+    "<extra_id_23>",
+    "<extra_id_24>",
+    "<extra_id_25>",
+    "<extra_id_26>",
+    "<extra_id_27>",
+    "<extra_id_28>",
+    "<extra_id_29>",
+    "<extra_id_30>",
+    "<extra_id_31>",
+    "<extra_id_32>",
+    "<extra_id_33>",
+    "<extra_id_34>",
+    "<extra_id_35>",
+    "<extra_id_36>",
+    "<extra_id_37>",
+    "<extra_id_38>",
+    "<extra_id_39>",
+    "<extra_id_40>",
+    "<extra_id_41>",
+    "<extra_id_42>",
+    "<extra_id_43>",
+    "<extra_id_44>",
+    "<extra_id_45>",
+    "<extra_id_46>",
+    "<extra_id_47>",
+    "<extra_id_48>",
+    "<extra_id_49>",
+    "<extra_id_50>",
+    "<extra_id_51>",
+    "<extra_id_52>",
+    "<extra_id_53>",
+    "<extra_id_54>",
+    "<extra_id_55>",
+    "<extra_id_56>",
+    "<extra_id_57>",
+    "<extra_id_58>",
+    "<extra_id_59>",
+    "<extra_id_60>",
+    "<extra_id_61>",
+    "<extra_id_62>",
+    "<extra_id_63>",
+    "<extra_id_64>",
+    "<extra_id_65>",
+    "<extra_id_66>",
+    "<extra_id_67>",
+    "<extra_id_68>",
+    "<extra_id_69>",
+    "<extra_id_70>",
+    "<extra_id_71>",
+    "<extra_id_72>",
+    "<extra_id_73>",
+    "<extra_id_74>",
+    "<extra_id_75>",
+    "<extra_id_76>",
+    "<extra_id_77>",
+    "<extra_id_78>",
+    "<extra_id_79>",
+    "<extra_id_80>",
+    "<extra_id_81>",
+    "<extra_id_82>",
+    "<extra_id_83>",
+    "<extra_id_84>",
+    "<extra_id_85>",
+    "<extra_id_86>",
+    "<extra_id_87>",
+    "<extra_id_88>",
+    "<extra_id_89>",
+    "<extra_id_90>",
+    "<extra_id_91>",
+    "<extra_id_92>",
+    "<extra_id_93>",
+    "<extra_id_94>",
+    "<extra_id_95>",
+    "<extra_id_96>",
+    "<extra_id_97>",
+    "<extra_id_98>",
+    "<extra_id_99>",
+    "<extra_id_100>",
+    "<extra_id_101>",
+    "<extra_id_102>",
+    "<extra_id_103>",
+    "<extra_id_104>",
+    "<extra_id_105>",
+    "<extra_id_106>",
+    "<extra_id_107>",
+    "<extra_id_108>",
+    "<extra_id_109>",
+    "<extra_id_110>",
+    "<extra_id_111>",
+    "<extra_id_112>",
+    "<extra_id_113>",
+    "<extra_id_114>",
+    "<extra_id_115>",
+    "<extra_id_116>",
+    "<extra_id_117>",
+    "<extra_id_118>",
+    "<extra_id_119>",
+    "<extra_id_120>",
+    "<extra_id_121>",
+    "<extra_id_122>",
+    "<extra_id_123>",
+    "<extra_id_124>"
+  ],
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer_config.json

@@ -0,0 +1,157 @@
+{
+  "additional_special_tokens": [
+    "<extra_id_0>",
+    "<extra_id_1>",
+    "<extra_id_2>",
+    "<extra_id_3>",
+    "<extra_id_4>",
+    "<extra_id_5>",
+    "<extra_id_6>",
+    "<extra_id_7>",
+    "<extra_id_8>",
+    "<extra_id_9>",
+    "<extra_id_10>",
+    "<extra_id_11>",
+    "<extra_id_12>",
+    "<extra_id_13>",
+    "<extra_id_14>",
+    "<extra_id_15>",
+    "<extra_id_16>",
+    "<extra_id_17>",
+    "<extra_id_18>",
+    "<extra_id_19>",
+    "<extra_id_20>",
+    "<extra_id_21>",
+    "<extra_id_22>",
+    "<extra_id_23>",
+    "<extra_id_24>",
+    "<extra_id_25>",
+    "<extra_id_26>",
+    "<extra_id_27>",
+    "<extra_id_28>",
+    "<extra_id_29>",
+    "<extra_id_30>",
+    "<extra_id_31>",
+    "<extra_id_32>",
+    "<extra_id_33>",
+    "<extra_id_34>",
+    "<extra_id_35>",
+    "<extra_id_36>",
+    "<extra_id_37>",
+    "<extra_id_38>",
+    "<extra_id_39>",
+    "<extra_id_40>",
+    "<extra_id_41>",
+    "<extra_id_42>",
+    "<extra_id_43>",
+    "<extra_id_44>",
+    "<extra_id_45>",
+    "<extra_id_46>",
+    "<extra_id_47>",
+    "<extra_id_48>",
+    "<extra_id_49>",
+    "<extra_id_50>",
+    "<extra_id_51>",
+    "<extra_id_52>",
+    "<extra_id_53>",
+    "<extra_id_54>",
+    "<extra_id_55>",
+    "<extra_id_56>",
+    "<extra_id_57>",
+    "<extra_id_58>",
+    "<extra_id_59>",
+    "<extra_id_60>",
+    "<extra_id_61>",
+    "<extra_id_62>",
+    "<extra_id_63>",
+    "<extra_id_64>",
+    "<extra_id_65>",
+    "<extra_id_66>",
+    "<extra_id_67>",
+    "<extra_id_68>",
+    "<extra_id_69>",
+    "<extra_id_70>",
+    "<extra_id_71>",
+    "<extra_id_72>",
+    "<extra_id_73>",
+    "<extra_id_74>",
+    "<extra_id_75>",
+    "<extra_id_76>",
+    "<extra_id_77>",
+    "<extra_id_78>",
+    "<extra_id_79>",
+    "<extra_id_80>",
+    "<extra_id_81>",
+    "<extra_id_82>",
+    "<extra_id_83>",
+    "<extra_id_84>",
+    "<extra_id_85>",
+    "<extra_id_86>",
+    "<extra_id_87>",
+    "<extra_id_88>",
+    "<extra_id_89>",
+    "<extra_id_90>",
+    "<extra_id_91>",
+    "<extra_id_92>",
+    "<extra_id_93>",
+    "<extra_id_94>",
+    "<extra_id_95>",
+    "<extra_id_96>",
+    "<extra_id_97>",
+    "<extra_id_98>",
+    "<extra_id_99>",
+    "<extra_id_100>",
+    "<extra_id_101>",
+    "<extra_id_102>",
+    "<extra_id_103>",
+    "<extra_id_104>",
+    "<extra_id_105>",
+    "<extra_id_106>",
+    "<extra_id_107>",
+    "<extra_id_108>",
+    "<extra_id_109>",
+    "<extra_id_110>",
+    "<extra_id_111>",
+    "<extra_id_112>",
+    "<extra_id_113>",
+    "<extra_id_114>",
+    "<extra_id_115>",
+    "<extra_id_116>",
+    "<extra_id_117>",
+    "<extra_id_118>",
+    "<extra_id_119>",
+    "<extra_id_120>",
+    "<extra_id_121>",
+    "<extra_id_122>",
+    "<extra_id_123>",
+    "<extra_id_124>"
+  ],
+  "clean_up_tokenization_spaces": true,
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "extra_ids": 125,
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": {
+    "__type": "AddedToken",
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "tokenizer_class": "ByT5Tokenizer",
+  "unk_token": {
+    "__type": "AddedToken",
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}