modeling_wilai.py

from typing import Tuple

import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import CrossEntropyLoss

from transformers.modeling_utils import PreTrainedModel
from transformers.generation import GenerationMixin
from transformers.modeling_outputs import CausalLMOutputWithPast, BaseModelOutputWithPast
from transformers.utils import logging
from transformers.utils import add_start_docstrings, add_start_docstrings_to_model_forward
from transformers.activations import ACT2FN
from .configuration_wilai import WilaiConfig


logger = logging.get_logger(__name__)

_TOKENIZER_FOR_DOC = "WilaiTokenizer"
_CHECKPOINT_FOR_DOC = "JonusNattapong/wilai-2.0"
_CONFIG_FOR_DOC = "WilaiConfig"


class WilaiAttention(nn.Module):
    def __init__(self, config, is_cross_attention=False, layer_idx=None):
        super().__init__()

        max_positions = config.max_position_embeddings
        self.register_buffer(
            "bias",
            torch.tril(torch.ones((max_positions, max_positions), dtype=torch.uint8)).view(
                1, 1, max_positions, max_positions
            ),
        )
        self.register_buffer("masked_bias", torch.tensor(-1e4))

        self.embed_dim = config.hidden_size
        self.num_attention_heads = config.num_attention_heads
        self.head_dim = self.embed_dim // self.num_attention_heads
        if self.head_dim * self.num_attention_heads != self.embed_dim:
            raise ValueError(
                f"embed_dim must be divisible by num_attention_heads (got `embed_dim`: {self.embed_dim} and"
                f" `num_attention_heads`: {self.num_attention_heads})."
            )
        self.scale_attn = torch.sqrt(torch.tensor(self.head_dim, dtype=torch.float32)).to(torch.get_default_dtype())

        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=config.attn_pdrop == 0.0)
        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=config.attn_pdrop == 0.0)
        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=config.attn_pdrop == 0.0)
        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=config.attn_pdrop == 0.0)
        self.dropout = config.attn_pdrop

        self.pruned_heads = set()
        self.layer_idx = layer_idx
        self.is_cross_attention = is_cross_attention

    def prune_heads(self, heads):
        if len(heads) == 0:
            return
        # For simplicity, we'll skip pruning for now
        # This would require implementing the prune functions
        pass

    def _attn(self, query, key, value, attention_mask=None, head_mask=None):
        # compute causal mask from causal mask buffer
        query_length, key_length = query.size(-2), key.size(-2)
        causal_mask = self.bias[:, :, key_length - query_length: key_length, :key_length].bool()

        # Keep the attention weights computation in fp32 to avoid overflow issues
        query = query.to(torch.float32)
        key = key.to(torch.float32)

        attn_weights = torch.matmul(query, key.transpose(-1, -2))

        attn_weights = attn_weights / self.scale_attn
        attn_weights = torch.where(causal_mask, attn_weights, self.masked_bias.to(attn_weights.dtype))

        if attention_mask is not None:
            # Apply the attention mask
            attn_weights = attn_weights + attention_mask

        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
        attn_weights = attn_weights.to(value.dtype)
        attn_weights = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)

        # Mask heads if we want to
        if head_mask is not None:
            attn_weights = attn_weights * head_mask

        attn_output = torch.matmul(attn_weights, value)

        return attn_output, attn_weights

    def forward(
        self,
        hidden_states,
        layer_past=None,
        attention_mask=None,
        head_mask=None,
        use_cache=False,
        output_attentions=False,
    ):
        query = self.q_proj(hidden_states)
        key = self.k_proj(hidden_states)
        value = self.v_proj(hidden_states)

        query = self._split_heads(query, self.num_attention_heads, self.head_dim)
        key = self._split_heads(key, self.num_attention_heads, self.head_dim)
        value = self._split_heads(value, self.num_attention_heads, self.head_dim)

        if layer_past is not None:
            past_key, past_value = layer_past
            key = torch.cat((past_key, key), dim=-2)
            value = torch.cat((past_value, value), dim=-2)

        if use_cache is True:
            present = (key, value)
        else:
            present = None

        attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask)

        attn_output = self._merge_heads(attn_output, self.num_attention_heads, self.head_dim)
        attn_output = self.out_proj(attn_output)
        attn_output = nn.functional.dropout(attn_output, p=self.dropout, training=self.training)

        outputs = (attn_output, present)
        if output_attentions:
            outputs += (attn_weights,)

        return outputs  # a, present, (attentions)

    def _split_heads(self, tensor, num_attention_heads, attn_head_size):
        """
        Splits hidden_size dim into attn_head_size and num_attention_heads
        """
        new_shape = tensor.size()[:-1] + (num_attention_heads, attn_head_size)
        tensor = tensor.view(*new_shape)
        return tensor.permute(0, 2, 1, 3)  # (batch, head, seq_length, head_features)

    def _merge_heads(self, tensor, num_attention_heads, attn_head_size):
        """
        Merges attn_head_size dim and num_attn_heads dim into hidden_size
        """
        tensor = tensor.permute(0, 2, 1, 3).contiguous()
        new_shape = tensor.size()[:-2] + (num_attention_heads * attn_head_size,)
        return tensor.view(new_shape)


class WilaiMLP(nn.Module):
    def __init__(self, intermediate_size, config):  # in MLP: intermediate_size= 4 * hidden_size
        super().__init__()
        embed_dim = config.hidden_size

        self.c_fc = nn.Linear(embed_dim, intermediate_size)
        self.c_proj = nn.Linear(intermediate_size, embed_dim)
        self.act = ACT2FN[config.activation_function]
        self.dropout = nn.Dropout(config.resid_pdrop)

    def forward(self, hidden_states):
        hidden_states = self.c_fc(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states = self.c_proj(hidden_states)
        hidden_states = self.dropout(hidden_states)
        return hidden_states


class WilaiBlock(nn.Module):
    def __init__(self, config, layer_idx=None):
        super().__init__()
        inner_dim = config.n_inner if config.n_inner is not None else 4 * config.hidden_size
        self.ln_1 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_epsilon)
        self.attn = WilaiAttention(config, layer_idx=layer_idx)
        self.ln_2 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_epsilon)
        self.mlp = WilaiMLP(inner_dim, config)

    def forward(
        self,
        hidden_states,
        layer_past=None,
        attention_mask=None,
        head_mask=None,
        use_cache=False,
        output_attentions=False,
    ):
        residual = hidden_states
        hidden_states = self.ln_1(hidden_states)
        attn_outputs = self.attn(
            hidden_states,
            layer_past=layer_past,
            attention_mask=attention_mask,
            head_mask=head_mask,
            use_cache=use_cache,
            output_attentions=output_attentions,
        )
        attn_output = attn_outputs[0]  # output_attn: a, present, (attentions)
        outputs = attn_outputs[1:]
        # residual connection
        hidden_states = attn_output + residual

        residual = hidden_states
        hidden_states = self.ln_2(hidden_states)
        feed_forward_hidden_states = self.mlp(hidden_states)
        # residual connection
        hidden_states = residual + feed_forward_hidden_states

        if use_cache:
            outputs = (hidden_states,) + outputs
        else:
            outputs = (hidden_states,) + outputs[1:]

        return outputs  # hidden_states, present, (attentions)


class WilaiPreTrainedModel(PreTrainedModel):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and
    loading pretrained models.
    """

    config_class = WilaiConfig
    base_model_prefix = "transformer"
    _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.masked_bias", r"h\.\d+\.attn\.bias"]

    def __init__(self, *inputs, **kwargs):
        super().__init__(*inputs, **kwargs)

    def _init_weights(self, module):
        """Initialize the weights."""
        if isinstance(module, (nn.Linear, nn.Embedding)):
            # Slightly different from Mesh Transformer JAX which uses truncated_normal for initialization
            # cf https://github.com/pytorch/pytorch/pull/5617
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if isinstance(module, nn.Linear) and module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)


WILAI_START_DOCSTRING = r"""

    This model inherits from :class:`~transformers.PreTrainedModel`. Check the superclass documentation for the generic
    methods the library implements for all its model (such as downloading or saving, resizing the input embeddings,
    pruning heads etc.)

    This model is also a PyTorch `torch.nn.Module <https://pytorch.org/docs/stable/nn.html#torch.nn.Module>`__
    subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to
    general usage and behavior.

    Parameters:
        config (:class:`~transformers.WilaiConfig`): Model configuration class with all the parameters of the model.
            Initializing with a config file does not load the weights associated with the model, only the
            configuration. Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model
            weights.
"""

WILAI_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`):
            :obj:`input_ids_length` = ``sequence_length`` if :obj:`past_key_values` is ``None`` else
            ``past_key_values[0][0].shape[-2]`` (``sequence_length`` of input past key value states). Indices of input
            sequence tokens in the vocabulary.

            If :obj:`past_key_values` is used, only ``input_ids`` that do not have their past calculated should be
            passed as ``input_ids``.

            Indices can be obtained using :class:`~transformers.WilaiTokenizer`. See
            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
            details.

            `What are input IDs? <../glossary.html#input-ids>`__
        past_key_values (:obj:`Tuple[Tuple[torch.Tensor]]` of length :obj:`config.n_layers`):
            Contains precomputed hidden-states (key and values in the attention blocks) as computed by the model (see
            :obj:`past_key_values` output below). Can be used to speed up sequential decoding. The ``input_ids`` which
            have their past given to this model should not be passed as ``input_ids`` as they have already been
            computed.
        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
            Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:

            - 1 for tokens that are **not masked**,
            - 0 for tokens that are **masked**.

            `What are attention masks? <../glossary.html#attention-mask>`__
        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`, `optional`):
            Segment token indices to indicate first and second portions of the inputs. Indices are selected in ``[0,
            1]``:

            - 0 corresponds to a `sentence A` token,
            - 1 corresponds to a `sentence B` token.

            `What are token type IDs? <../glossary.html#token-type-ids>`_
        position_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
            Indices of positions of each input sequence token in the position embeddings. Selected in the range ``[0,
            config.max_position_embeddings - 1]``.

            `What are position IDs? <../glossary.html#position-ids>`_
        head_mask (:obj:`torch.FloatTensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`):
            Mask to nullify selected heads of the self-attention modules. Mask values selected in ``[0, 1]``:

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.

        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
            Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
            This is useful if you want more control over how to convert :obj:`input_ids` indices into associated
            vectors than the model's internal embedding lookup matrix.
        use_cache (:obj:`bool`, `optional`):
            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
            decoding (see :obj:`past_key_values`).
        output_attentions (:obj:`bool`, `optional`):
            Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
            tensors for more detail.
        output_hidden_states (:obj:`bool`, `optional`):
            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
            more detail.
        return_dict (:obj:`bool`, `optional`):
            Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
"""


@add_start_docstrings(
    "The bare Wilai Model transformer outputting raw hidden-states without any specific head on top.",
    WILAI_START_DOCSTRING,
)
class WilaiModel(WilaiPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.embed_dim = config.hidden_size
        self.vocab_size = config.vocab_size
        self.wte = nn.Embedding(config.vocab_size, self.embed_dim)
        self.wpe = nn.Embedding(config.max_position_embeddings, self.embed_dim)
        self.drop = nn.Dropout(config.embd_pdrop)
        self.h = nn.ModuleList([WilaiBlock(config, layer_idx=i) for i in range(config.num_hidden_layers)])
        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)

        self.init_weights()

    def get_input_embeddings(self):
        return self.wte

    def set_input_embeddings(self, new_embeddings):
        self.wte = new_embeddings

    def get_head_mask(self, head_mask, num_hidden_layers):
        """Simple head mask implementation."""
        if head_mask is not None:
            return head_mask
        return [None] * num_hidden_layers

    @add_start_docstrings_to_model_forward(WILAI_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids=None,
        past_key_values=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        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
        )
        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

        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both input_ids and 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])
            batch_size = input_ids.shape[0]
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
            batch_size = inputs_embeds.shape[0]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        device = input_ids.device if input_ids is not None else inputs_embeds.device

        if token_type_ids is not None:
            token_type_ids = token_type_ids.view(-1, input_shape[-1])

        if past_key_values is None:
            past_length = 0
            past_key_values = tuple([None] * len(self.h))
        else:
            past_length = past_key_values[0][0].size(-2)

        if position_ids is None:
            position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
            position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])

        # Attention mask.
        if attention_mask is not None:
            assert batch_size > 0, "batch_size has to be defined and > 0"
            attention_mask = attention_mask.view(batch_size, -1)
            # We create a 3D attention mask from a 2D tensor mask.
            # Sizes are [batch_size, 1, 1, to_seq_length]
            # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
            # this attention mask is more simple than the triangular masking of causal attention
            # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
            attention_mask = attention_mask[:, None, None, :]

            # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
            # masked positions, this operation will create a tensor which is 0.0 for
            # positions we want to attend and -10000.0 for masked positions.
            # Since we're adding it to the raw scores before the softmax, this is
            # effectively the same as removing these entirely.
            attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility
            attention_mask = (1.0 - attention_mask) * -10000.0

        # Prepare head mask if needed
        # 1.0 in head_mask indicate we keep the head
        # attention_probs has shape bsz x n_heads x N x N
        # head_mask has shape n_layer x n_head or head_mask has shape n_layer x batch x n_head x N x N
        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)

        if inputs_embeds is None:
            inputs_embeds = self.wte(input_ids)
        position_embeds = self.wpe(position_ids)
        hidden_states = inputs_embeds + position_embeds

        if token_type_ids is not None:
            token_type_embeds = self.wte(token_type_ids)
            hidden_states = hidden_states + token_type_embeds

        hidden_states = self.drop(hidden_states)

        output_shape = input_shape + (hidden_states.size(-1),)

        presents = () if use_cache else None
        all_self_attentions = () if output_attentions else None
        all_hidden_states = () if output_hidden_states else None

        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            # Use the configuration flag for gradient checkpointing
            if getattr(self.config, "gradient_checkpointing", False) and self.training:

                if use_cache:
                    logger.warning(
                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                    )
                    use_cache = False

                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        # None for past_key_value
                        return module(*inputs, use_cache, output_attentions)

                    return custom_forward

                outputs = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(block),
                    hidden_states,
                    None,
                    attention_mask,
                    head_mask[i],
                )
            else:
                outputs = block(
                    hidden_states,
                    layer_past=layer_past,
                    attention_mask=attention_mask,
                    head_mask=head_mask[i],
                    use_cache=use_cache,
                    output_attentions=output_attentions,
                )

            hidden_states = outputs[0]
            if use_cache is True:
                presents = presents + (outputs[1],)

            if output_attentions:
                all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)

        hidden_states = self.ln_f(hidden_states)
        hidden_states = hidden_states.view(*output_shape)

        # Add last hidden state
        if output_hidden_states:
            all_hidden_states = all_hidden_states + (hidden_states,)

        if not return_dict:
            return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)

        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=presents,
            hidden_states=all_hidden_states,
            attentions=all_self_attentions,
        )


@add_start_docstrings(
    """
    The Wilai Model transformer with a language modeling head on top (linear layer with weights tied to the input
    embeddings).
    """,
    WILAI_START_DOCSTRING,
)
class WilaiForCausalLM(WilaiPreTrainedModel, GenerationMixin):
    _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.masked_bias", r"h\.\d+\.attn\.bias", r"lm_head\.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.transformer = WilaiModel(config)
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

        self.init_weights()

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):
        token_type_ids = kwargs.get("token_type_ids", None)
        # only last token for inputs_ids if past is defined in kwargs
        if past:
            input_ids = input_ids[:, -1].unsqueeze(-1)
            if token_type_ids is not None:
                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)

        attention_mask = kwargs.get("attention_mask", None)
        position_ids = kwargs.get("position_ids", None)

        if attention_mask is not None and position_ids is None:
            # create position_ids on the fly for batch generation
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)
            if past:
                position_ids = position_ids[:, -1].unsqueeze(-1)
        else:
            position_ids = None
        return {
            "input_ids": input_ids,
            "past_key_values": past,
            "use_cache": kwargs.get("use_cache"),
            "position_ids": position_ids,
            "attention_mask": attention_mask,
            "token_type_ids": token_type_ids,
        }

    @add_start_docstrings_to_model_forward(WILAI_INPUTS_DOCSTRING)
    def forward(
        self,
        input_ids=None,
        past_key_values=None,
        attention_mask=None,
        token_type_ids=None,
        position_ids=None,
        head_mask=None,
        inputs_embeds=None,
        labels=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        r"""
        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):
            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
            ``labels = input_ids`` Indices are selected in ``[-100, 0, ..., config.vocab_size]`` All labels set to
            ``-100`` are ignored (masked), the loss is only computed for labels in ``[0, ..., config.vocab_size]``
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        transformer_outputs = self.transformer(
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        hidden_states = transformer_outputs.last_hidden_state if return_dict else transformer_outputs[0]

        lm_logits = self.lm_head(hidden_states)

        loss = None
        if labels is not None:
            # Shift so that tokens < n predict n
            shift_logits = lm_logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))

        if not return_dict:
            output = (lm_logits,) + transformer_outputs[1:]
            return ((loss,) + output) if loss is not None else output

        return CausalLMOutputWithPast(
            loss=loss,
            logits=lm_logits,
            past_key_values=transformer_outputs.past_key_values if return_dict else transformer_outputs[1],
            hidden_states=transformer_outputs.hidden_states if return_dict else transformer_outputs[2],
            attentions=transformer_outputs.attentions if return_dict else transformer_outputs[3],
        )

    @staticmethod
    def _reorder_cache(past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor) -> Tuple[Tuple[torch.Tensor]]:
        """
        This function is used to re-order the :obj:`past_key_values` cache if
        :meth:`~transformers.PreTrainedModel.beam_search` or :meth:`~transformers.PreTrainedModel.beam_sample` is
        called. This is required to match :obj:`past_key_values` with the correct beam_idx at every generation step.
        """
        return tuple(
            tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)
            for layer_past in past
        )

    def get_input_embeddings(self):
        return self.transformer.wte

    def set_input_embeddings(self, new_embeddings):
        self.transformer.wte = new_embeddings