source/official-code/micro_hf/models/hard_alibi_model.py

""" """

from typing import Optional, Tuple, Union

import torch
import torch.utils.checkpoint
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss

from transformers.activations import ACT2FN
from transformers.modeling_outputs import (
    BaseModelOutputWithPast,
    CausalLMOutputWithPast,
    QuestionAnsweringModelOutput,
    SequenceClassifierOutputWithPast,
    TokenClassifierOutput,
)
from transformers.modeling_utils import PreTrainedModel
from transformers.utils import logging

from transformers.models.gpt_neox.configuration_gpt_neox import GPTNeoXConfig

from transformers.models.gpt_neox.modeling_gpt_neox import GPTNeoXPreTrainedModel, GPTNeoXMLP

logger = logging.get_logger(__name__)




class GPTHardAlibiAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.num_attention_heads = config.num_attention_heads
        self.hidden_size = config.hidden_size
        if self.hidden_size % self.num_attention_heads != 0:
            raise ValueError(
                "The hidden size is not divisble by the number of attention heads! Make sure to update them"
            )
        self.head_size = self.hidden_size // self.num_attention_heads
        self.rotary_ndims = int(self.head_size * config.rotary_pct)
        self._init_bias(config.max_position_embeddings)

        self.register_buffer("masked_bias", torch.tensor(-1e9), persistent=False)
        self.norm_factor = self.head_size**-0.5
        self.query_key_value = nn.Linear(config.hidden_size, 3 * config.hidden_size)
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
        self.attention_dropout = nn.Dropout(config.attention_dropout)

    def _init_bias(self, max_positions, device=None):
        self.register_buffer(
            "bias",
            torch.tril(torch.ones((max_positions, max_positions), dtype=torch.bool)).view(
                1, 1, max_positions, max_positions
            ),
            persistent=False,
        )
        if device is not None:
            self.bias = self.bias.to(device)

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        attention_mask: torch.FloatTensor,
        position_ids: torch.LongTensor,
        head_mask: Optional[torch.FloatTensor] = None,
        layer_past: Optional[Tuple[torch.Tensor]] = None,
        use_cache: Optional[bool] = False,
        output_attentions: Optional[bool] = False,
    ):
        has_layer_past = layer_past is not None

        # Compute QKV
        # Attention heads [batch, seq_len, hidden_size]
        #   --> [batch, seq_len, (np * 3 * head_size)]
        qkv = self.query_key_value(hidden_states)

        # [batch, seq_len, (num_heads * 3 * head_size)]
        #   --> [batch, seq_len, num_heads, 3 * head_size]
        new_qkv_shape = qkv.size()[:-1] + (self.num_attention_heads, 3 * self.head_size)
        qkv = qkv.view(*new_qkv_shape)

        # [batch, seq_len, num_attention_heads, 3 * head_size] --> 3 [batch, num_attention_heads, seq_len, head_size]
        query = qkv[..., : self.head_size].permute(0, 2, 1, 3)
        key = qkv[..., self.head_size : 2 * self.head_size].permute(0, 2, 1, 3)
        value = qkv[..., 2 * self.head_size :].permute(0, 2, 1, 3)

        # Compute rotary embeddings on rotary_ndims
        query_rot = query[..., : self.rotary_ndims]
        query_pass = query[..., self.rotary_ndims :]
        key_rot = key[..., : self.rotary_ndims]
        key_pass = key[..., self.rotary_ndims :]

        # Compute token offset for rotary embeddings (when decoding)
        seq_len = key.shape[-2]
        if has_layer_past:
            seq_len += layer_past[0].shape[-2]
        query = torch.cat((query, query_pass), dim=-1)
        key = torch.cat((key, key_pass), dim=-1)

        # Cache QKV values
        if has_layer_past:
            past_key = layer_past[0]
            past_value = layer_past[1]
            key = torch.cat((past_key, key), dim=-2)
            value = torch.cat((past_value, value), dim=-2)
        present = (key, value) if use_cache else None

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

        # Reshape outputs
        attn_output = self._merge_heads(attn_output, self.num_attention_heads, self.head_size)
        attn_output = self.dense(attn_output)

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

        return outputs

    @classmethod
    def _split_heads(cls, tensor, num_attention_heads, attn_head_size):
        """
        Splits hidden dim into attn_head_size and num_attention_heads
        """
        # tensor: [bs, seq_len, hidden_size]
        new_shape = tensor.size()[:-1] + (num_attention_heads, attn_head_size)
        # -> [bs, seq_len, num_attention_heads, attn_head_size]
        tensor = tensor.view(new_shape)
        # -> [bs, num_attention_heads, seq_len, attn_head_size]
        tensor = tensor.permute(0, 2, 1, 3)
        return tensor

    @classmethod
    def _merge_heads(cls, tensor, num_attention_heads, attn_head_size):
        """
        Merges attn_head_size dim and num_attn_heads dim into hidden dim
        """
        # tensor [bs, num_attention_heads, seq_len, attn_head_size]
        tensor = tensor.permute(0, 2, 1, 3).contiguous()
        # -> [bs, seq_len, num_attention_heads, attn_head_size]
        tensor = tensor.view(tensor.size(0), tensor.size(1), num_attention_heads * attn_head_size)
        # -> [bs, seq_len, hidden_size]
        return tensor

    def _attn(self, query, key, value, attention_mask=None, head_mask=None):
        # q, k, v: [bs, num_attention_heads, seq_len, attn_head_size]
        # compute causal mask from causal mask buffer
        batch_size, num_attention_heads, query_length, attn_head_size = query.size()
        key_length = key.size(-2)

        # dynamically increase the causal mask with the key length, if needed.
        if key_length > self.bias.shape[-1]:
            self._init_bias(key_length, device=key.device)
        causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length]

        query = query.view(batch_size * num_attention_heads, query_length, attn_head_size)
        key = key.view(batch_size * num_attention_heads, key_length, attn_head_size)
        attn_scores = torch.zeros(
            batch_size * num_attention_heads,
            query_length,
            key_length,
            dtype=query.dtype,
            device=key.device,
        )
        attn_scores = torch.baddbmm(
            attn_scores,
            query,
            key.transpose(1, 2),
            beta=1.0,
            alpha=self.norm_factor,
        )
        attn_scores = attn_scores.view(batch_size, num_attention_heads, query_length, key_length)

        mask_value = torch.finfo(attn_scores.dtype).min
        # Need to be a tensor, otherwise we get error: `RuntimeError: expected scalar type float but found double`.
        # Need to be on the same device, otherwise `RuntimeError: ..., x and y to be on the same device`
        mask_value = torch.tensor(mask_value, dtype=attn_scores.dtype).to(attn_scores.device)
        attn_scores = torch.where(causal_mask, attn_scores, mask_value)
        
        if attention_mask is not None:
            # Apply the attention mask
            attn_scores = attn_scores + attention_mask
        
        
        #masking
        new_attn_scores = attn_scores.clone()
        for i in range(self.config.num_masked_heads):
            mask_head =  torch.ones_like(attn_scores[:,0]) * mask_value
            for j in range(i+1):
                mask_head[:,range(j,mask_head.shape[1]), range(0,mask_head.shape[2]-j)] = 0 
            new_attn_scores[:,i] = attn_scores[:,i] + mask_head
            
        attn_weights = nn.functional.softmax(new_attn_scores, dim=-1)
        attn_weights = attn_weights.to(value.dtype)

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

        attn_weights = self.attention_dropout(attn_weights)

        attn_output = torch.matmul(attn_weights, value)
        return attn_output, attn_weights



class GPTNeoXHardAlibiLayer(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.use_parallel_residual = config.use_parallel_residual
        self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.post_attention_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.post_attention_dropout = nn.Dropout(config.hidden_dropout)
        self.post_mlp_dropout = nn.Dropout(config.hidden_dropout)
        self.attention = GPTHardAlibiAttention(config)
        self.mlp = GPTNeoXMLP(config)

    def forward(
        self,
        hidden_states: Optional[torch.FloatTensor],
        attention_mask: Optional[torch.FloatTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = False,
        layer_past: Optional[Tuple[torch.Tensor]] = None,
        output_attentions: Optional[bool] = False,
    ):
        attention_layer_outputs = self.attention(
            self.input_layernorm(hidden_states),
            attention_mask=attention_mask,
            position_ids=position_ids,
            layer_past=layer_past,
            head_mask=head_mask,
            use_cache=use_cache,
            output_attentions=output_attentions,
        )
        attn_output = attention_layer_outputs[0]  # output_attn: attn_output, present, (attn_weights)
        attn_output = self.post_attention_dropout(attn_output)
        outputs = attention_layer_outputs[1:]

        if self.use_parallel_residual:
            # pseudocode:
            # x = x + attn(ln1(x)) + mlp(ln2(x))
            mlp_output = self.mlp(self.post_attention_layernorm(hidden_states))
            mlp_output = self.post_mlp_dropout(mlp_output)
            hidden_states = mlp_output + attn_output + hidden_states
        else:
            # pseudocode:
            # x = x + attn(ln1(x))
            # x = x + mlp(ln2(x))
            attn_output = attn_output + hidden_states
            mlp_output = self.mlp(self.post_attention_layernorm(attn_output))
            mlp_output = self.post_mlp_dropout(mlp_output)
            hidden_states = mlp_output + attn_output

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

        return outputs



class GPTNeoXHardAlibiModel(GPTNeoXPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        self.config = config

        self.embed_in = nn.Embedding(config.vocab_size, config.hidden_size)
        self.emb_dropout = nn.Dropout(config.hidden_dropout)
        self.layers = nn.ModuleList([GPTNeoXHardAlibiLayer(config) for _ in range(config.num_hidden_layers)])
        self.final_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)

        self.gradient_checkpointing = False

        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        return self.embed_in

    def set_input_embeddings(self, value):
        self.embed_in = value
    
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, BaseModelOutputWithPast]:
        r"""
        past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
        use_cache (`bool`, *optional*):
            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
            `past_key_values`).
        """
        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
        use_cache = use_cache if use_cache is not None else self.config.use_cache

        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:
            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
            input_shape = input_ids.size()
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        batch_size, seq_length = input_shape

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

        if position_ids is None:
            device = input_ids.device if input_ids is not None else inputs_embeds.device
            position_ids = torch.arange(past_length, seq_length + past_length, dtype=torch.long, device=device)
            position_ids = position_ids.unsqueeze(0)

        # 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 the dtype's smallest value for masked positions.
            # Since we are 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) * torch.finfo(self.dtype).min

        # 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
        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)

        if inputs_embeds is None:
            inputs_embeds = self.embed_in(input_ids)

        hidden_states = self.emb_dropout(inputs_embeds)

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

        presents = () if use_cache else None
        all_attentions = () if output_attentions else None
        all_hidden_states = () if output_hidden_states else None
        for i, (layer, layer_past) in enumerate(zip(self.layers, past_key_values)):
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            if self.gradient_checkpointing and self.training:
                outputs = self._gradient_checkpointing_func(
                    layer.__call__,
                    hidden_states,
                    attention_mask,
                    position_ids,
                    head_mask[i],
                    use_cache,
                    None,
                    output_attentions,
                )
            else:
                outputs = layer(
                    hidden_states,
                    attention_mask=attention_mask,
                    position_ids=position_ids,
                    head_mask=head_mask[i],
                    layer_past=layer_past,
                    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_attentions = all_attentions + (outputs[2 if use_cache else 1],)

        hidden_states = self.final_layer_norm(hidden_states)
        # 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_attentions] if v is not None)

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


class GPTNeoXHardAlibiForCausalLM(GPTNeoXPreTrainedModel):
    _tied_weights_keys = ["embed_out.weight"]

    def __init__(self, config):
        super().__init__(config)

        self.gpt_neox = GPTNeoXHardAlibiModel(config)
        self.embed_out = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

        # Initialize weights and apply final processing
        self.post_init()

    def get_output_embeddings(self):
        return self.embed_out

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

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        past_key_values: Optional[Tuple[Tuple[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, CausalLMOutputWithPast]:
        r"""
        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`. The two additional tensors are
            only required when the model is used as a decoder in a Sequence to Sequence model.

            Contains pre-computed hidden-states (key and values in the self-attention blocks that can be used (see
            `past_key_values` input) to speed up sequential decoding.

            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in
            `[-100, 0, ..., config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are
            ignored (masked), the loss is only computed for the tokens with labels n `[0, ..., config.vocab_size]`.
        use_cache (`bool`, *optional*):
            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
            `past_key_values`).

        Returns:

        Example:

        ```python
        >>> from transformers import AutoTokenizer, GPTNeoXForCausalLM, GPTNeoXConfig
        >>> import torch

        >>> tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-neox-20b")
        >>> config = GPTNeoXConfig.from_pretrained("EleutherAI/gpt-neox-20b")
        >>> config.is_decoder = True
        >>> model = GPTNeoXForCausalLM.from_pretrained("EleutherAI/gpt-neox-20b", config=config)

        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
        >>> outputs = model(**inputs)

        >>> prediction_logits = outputs.logits
        ```"""
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = self.gpt_neox(
            input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            past_key_values=past_key_values,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        hidden_states = outputs[0]
        lm_logits = self.embed_out(hidden_states)

        lm_loss = None
        if labels is not None:
            # move labels to correct device to enable model parallelism
            labels = labels.to(lm_logits.device)
            # we are doing next-token prediction; shift prediction scores and input ids by one
            shift_logits = lm_logits[:, :-1, :].contiguous()
            labels = labels[:, 1:].contiguous()
            loss_fct = CrossEntropyLoss()
            lm_loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), labels.view(-1))

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

        return CausalLMOutputWithPast(
            loss=lm_loss,
            logits=lm_logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

    def prepare_inputs_for_generation(
        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
    ):
        input_shape = input_ids.shape
        # cut decoder_input_ids if past is used
        if past_key_values is not None:
            past_length = past_key_values[0][0].shape[2]

            # Some generation methods already pass only the last input ID
            if input_ids.shape[1] > past_length:
                remove_prefix_length = past_length
            else:
                # Default to old behavior: keep only final ID
                remove_prefix_length = input_ids.shape[1] - 1

            input_ids = input_ids[:, remove_prefix_length:]

        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_key_values:
                position_ids = position_ids[:, -input_ids.shape[1] :]

        # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly
        if attention_mask is None:
            attention_mask = input_ids.new_ones(input_shape)

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and past_key_values is None:
            model_inputs = {"inputs_embeds": inputs_embeds}
        else:
            model_inputs = {"input_ids": input_ids}
        model_inputs.update(
            {
                "attention_mask": attention_mask,
                "past_key_values": past_key_values,
                "position_ids": position_ids,
            }
        )

        return model_inputs

    def _reorder_cache(self, past_key_values, beam_idx):
        reordered_past = ()
        for layer_past in past_key_values:
            reordered_past += (
                tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past[:2])
                + layer_past[2:],
            )
        return reordered_past