Spaces:

ProCreations
/

repro-expressivity-efficiency-hybrid-sequence

Running

repro-expressivity-efficiency-hybrid-sequence / source /official-code /mini /models /rope.py

Reproduction logbook (paper-82EJxJzG6r)

4ca4e4c verified 3 months ago

23.1 kB

""" PyTorch GPTNeoX model."""

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

from transformers.activations import ACT2FN
from transformers.file_utils import (
    add_code_sample_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    replace_return_docstrings,
)
from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
from transformers.modeling_utils import PreTrainedModel
from transformers.utils import logging
from transformers.models.gpt_neox.configuration_gpt_neox import GPTNeoXConfig

from transformers.generation import GenerationMixin


logger = logging.get_logger(__name__)

_CHECKPOINT_FOR_DOC = "gpt-neox-20b"
_CONFIG_FOR_DOC = "GPTNeoXConfig"
_TOKENIZER_FOR_DOC = "GPTNeoXTokenizerFast"


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

    config_class = GPTNeoXConfig
    base_model_prefix = "gpt_neox"
    supports_gradient_checkpointing = True
    _no_split_modules = ["GPTNeoXLayer"]

    def _init_weights(self, module):
        """Initialize the weights"""
        if isinstance(module, nn.Linear):
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, nn.LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)

    def _set_gradient_checkpointing(self, module, value=False):
        if isinstance(module, GPTNeoXModel):
            module.gradient_checkpointing = value


class GPTNeoXAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.num_attention_heads = config.num_attention_heads
        self.hidden_size = config.hidden_size
        self.head_size = self.hidden_size // self.num_attention_heads
        self.rotary_ndims = int(self.head_size * config.rotary_pct)
        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(-1e9))
        self.rotary_emb = RotaryEmbedding(self.rotary_ndims, base=config.rotary_emb_base)
        self.norm_factor = torch.sqrt(torch.tensor(self.head_size, dtype=torch.float32)).to(torch.get_default_dtype())
        self.query_key_value = nn.Linear(config.hidden_size, 3 * config.hidden_size)
        self.dense = nn.Linear(config.hidden_size, config.hidden_size)

    def forward(
        self,
        hidden_states,
        attention_mask,
        head_mask=None,
        layer_past=None,
        use_cache=False,
        output_attentions=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]
        offset = 0
        if has_layer_past:
            offset = layer_past[0].shape[-2]
            seq_len += offset
        cos, sin = self.rotary_emb(value, seq_len=seq_len)
        query, key = apply_rotary_pos_emb(query_rot, key_rot, cos, sin, offset=offset)
        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 = None if use_cache else (key, value)

        # 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)

        causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length].bool()

        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.einsum("bik,bjk->bij", query, key) / self.norm_factor
        attn_scores = attn_scores.view(batch_size, num_attention_heads, query_length, key_length)

        attn_scores = torch.where(causal_mask, attn_scores, self.masked_bias.to(attn_scores.dtype))

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

        attn_weights = nn.functional.softmax(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_output = torch.matmul(attn_weights, value)
        return attn_output, attn_weights


def attention_mask_func(attention_scores, ltor_mask):
    attention_scores.masked_fill_(~ltor_mask, -10000.0)
    return attention_scores


class RotaryEmbedding(torch.nn.Module):
    def __init__(self, dim, base=10000, device=None):
        super().__init__()
        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
        self.register_buffer("inv_freq", inv_freq)
        self.max_seq_len_cached = None
        self.cos_cached = None
        self.sin_cached = None

    def forward(self, x, seq_len=None):
        # x: [bs, num_attention_heads, seq_len, head_size]
        if self.max_seq_len_cached is None or (seq_len > self.max_seq_len_cached):
            self.max_seq_len_cached = seq_len
            t = torch.arange(self.max_seq_len_cached, device=x.device, dtype=self.inv_freq.dtype)
            freqs = torch.einsum("i,j->ij", t, self.inv_freq)
            # Different from paper, but it uses a different permutation in order to obtain the same calculation
            emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
            self.cos_cached = emb.cos()[None, None, :, :]
            self.sin_cached = emb.sin()[None, None, :, :]
        return self.cos_cached[:seq_len, ...], self.sin_cached[:seq_len, ...]


def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_pos_emb(q, k, cos, sin, offset: int = 0):
    cos = cos[..., offset : q.shape[-2] + offset, :]
    sin = sin[..., offset : q.shape[-2] + offset, :]
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


class GPTNeoXMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.dense_h_to_4h = nn.Linear(config.hidden_size, config.intermediate_size)
        self.dense_4h_to_h = nn.Linear(config.intermediate_size, config.hidden_size)
        self.act = ACT2FN[config.hidden_act]

    def forward(self, hidden_states):
        hidden_states = self.dense_h_to_4h(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states = self.dense_4h_to_h(hidden_states)
        return hidden_states


class GPTNeoXLayer(nn.Module):
    def __init__(self, config):
        super().__init__()
        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.attention = GPTNeoXAttention(config)
        self.mlp = GPTNeoXMLP(config)

    def forward(
        self,
        hidden_states,
        attention_mask=None,
        head_mask=None,
        use_cache=False,
        layer_past=None,
        output_attentions=False,
    ):
        residual = hidden_states
        ln_out = self.input_layernorm(hidden_states)
        attention_layer_outputs = self.attention(
            ln_out,
            attention_mask=attention_mask,
            layer_past=layer_past,
            head_mask=head_mask,
            use_cache=use_cache,
            output_attentions=output_attentions,
        )
        attn_output = attention_layer_outputs[0]  # output_attn: a, present, (attentions)
        outputs = attention_layer_outputs[1:]

        mlp_output = self.mlp(self.post_attention_layernorm(hidden_states))
        hidden_states = mlp_output + attn_output + residual

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

        return outputs  # hidden_states, present, (attentions)


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

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

        # 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=None,
        attention_mask=None,
        head_mask=None,
        inputs_embeds=None,
        past_key_values=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        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:
            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_key_values = tuple([None] * self.config.num_hidden_layers)

        # 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 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) * -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
        # 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 = inputs_embeds

        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,)
            outputs = layer(
                hidden_states,
                attention_mask=attention_mask,
                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 GPTNeoXForCausalLM(GPTNeoXPreTrainedModel, GenerationMixin):

    _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]

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

        self.gpt_neox = GPTNeoXModel(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=None,
        attention_mask=None,
        inputs_embeds=None,
        head_mask=None,
        past_key_values=None,
        labels=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        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 GPTNeoXTokenizer, GPTNeoXForCausalLM, GPTNeoXConfig
        >>> import torch
        >>> tokenizer = GPTNeoXTokenizer.from_pretrained("gpt-neox-20b")
        >>> config = GPTNeoXConfig.from_pretrained("gpt-neox-20b")
        >>> config.is_decoder = True
        >>> model = GPTNeoXForCausalLM.from_pretrained("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,
            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:
            # 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=None, attention_mask=None, **model_kwargs):
        input_shape = input_ids.shape

        # 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)

        # cut decoder_input_ids if past is used
        if past is not None:
            input_ids = input_ids[:, -1:]

        return {"input_ids": input_ids, "attention_mask": attention_mask, "past_key_values": past}

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

	""" PyTorch GPTNeoX model."""

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

	from transformers.activations import ACT2FN
	from transformers.file_utils import (
	add_code_sample_docstrings,
	add_start_docstrings,
	add_start_docstrings_to_model_forward,
	replace_return_docstrings,
	)
	from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
	from transformers.modeling_utils import PreTrainedModel
	from transformers.utils import logging
	from transformers.models.gpt_neox.configuration_gpt_neox import GPTNeoXConfig

	from transformers.generation import GenerationMixin


	logger = logging.get_logger(__name__)

	_CHECKPOINT_FOR_DOC = "gpt-neox-20b"
	_CONFIG_FOR_DOC = "GPTNeoXConfig"
	_TOKENIZER_FOR_DOC = "GPTNeoXTokenizerFast"


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

	config_class = GPTNeoXConfig
	base_model_prefix = "gpt_neox"
	supports_gradient_checkpointing = True
	_no_split_modules = ["GPTNeoXLayer"]

	def _init_weights(self, module):
	"""Initialize the weights"""
	if isinstance(module, nn.Linear):
	module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
	if module.bias is not None:
	module.bias.data.zero_()
	elif isinstance(module, nn.Embedding):
	module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
	if module.padding_idx is not None:
	module.weight.data[module.padding_idx].zero_()
	elif isinstance(module, nn.LayerNorm):
	module.bias.data.zero_()
	module.weight.data.fill_(1.0)

	def _set_gradient_checkpointing(self, module, value=False):
	if isinstance(module, GPTNeoXModel):
	module.gradient_checkpointing = value


	class GPTNeoXAttention(nn.Module):
	def __init__(self, config):
	super().__init__()
	self.num_attention_heads = config.num_attention_heads
	self.hidden_size = config.hidden_size
	self.head_size = self.hidden_size // self.num_attention_heads
	self.rotary_ndims = int(self.head_size * config.rotary_pct)
	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(-1e9))
	self.rotary_emb = RotaryEmbedding(self.rotary_ndims, base=config.rotary_emb_base)
	self.norm_factor = torch.sqrt(torch.tensor(self.head_size, dtype=torch.float32)).to(torch.get_default_dtype())
	self.query_key_value = nn.Linear(config.hidden_size, 3 * config.hidden_size)
	self.dense = nn.Linear(config.hidden_size, config.hidden_size)

	def forward(
	self,
	hidden_states,
	attention_mask,
	head_mask=None,
	layer_past=None,
	use_cache=False,
	output_attentions=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]
	offset = 0
	if has_layer_past:
	offset = layer_past[0].shape[-2]
	seq_len += offset
	cos, sin = self.rotary_emb(value, seq_len=seq_len)
	query, key = apply_rotary_pos_emb(query_rot, key_rot, cos, sin, offset=offset)
	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 = None if use_cache else (key, value)

	# 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)

	causal_mask = self.bias[:, :, key_length - query_length : key_length, :key_length].bool()

	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.einsum("bik,bjk->bij", query, key) / self.norm_factor
	attn_scores = attn_scores.view(batch_size, num_attention_heads, query_length, key_length)

	attn_scores = torch.where(causal_mask, attn_scores, self.masked_bias.to(attn_scores.dtype))

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

	attn_weights = nn.functional.softmax(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_output = torch.matmul(attn_weights, value)
	return attn_output, attn_weights


	def attention_mask_func(attention_scores, ltor_mask):
	attention_scores.masked_fill_(~ltor_mask, -10000.0)
	return attention_scores


	class RotaryEmbedding(torch.nn.Module):
	def __init__(self, dim, base=10000, device=None):
	super().__init__()
	inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
	self.register_buffer("inv_freq", inv_freq)
	self.max_seq_len_cached = None
	self.cos_cached = None
	self.sin_cached = None

	def forward(self, x, seq_len=None):
	# x: [bs, num_attention_heads, seq_len, head_size]
	if self.max_seq_len_cached is None or (seq_len > self.max_seq_len_cached):
	self.max_seq_len_cached = seq_len
	t = torch.arange(self.max_seq_len_cached, device=x.device, dtype=self.inv_freq.dtype)
	freqs = torch.einsum("i,j->ij", t, self.inv_freq)
	# Different from paper, but it uses a different permutation in order to obtain the same calculation
	emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
	self.cos_cached = emb.cos()[None, None, :, :]
	self.sin_cached = emb.sin()[None, None, :, :]
	return self.cos_cached[:seq_len, ...], self.sin_cached[:seq_len, ...]


	def rotate_half(x):
	"""Rotates half the hidden dims of the input."""
	x1 = x[..., : x.shape[-1] // 2]
	x2 = x[..., x.shape[-1] // 2 :]
	return torch.cat((-x2, x1), dim=-1)


	def apply_rotary_pos_emb(q, k, cos, sin, offset: int = 0):
	cos = cos[..., offset : q.shape[-2] + offset, :]
	sin = sin[..., offset : q.shape[-2] + offset, :]
	q_embed = (q * cos) + (rotate_half(q) * sin)
	k_embed = (k * cos) + (rotate_half(k) * sin)
	return q_embed, k_embed


	class GPTNeoXMLP(nn.Module):
	def __init__(self, config):
	super().__init__()
	self.dense_h_to_4h = nn.Linear(config.hidden_size, config.intermediate_size)
	self.dense_4h_to_h = nn.Linear(config.intermediate_size, config.hidden_size)
	self.act = ACT2FN[config.hidden_act]

	def forward(self, hidden_states):
	hidden_states = self.dense_h_to_4h(hidden_states)
	hidden_states = self.act(hidden_states)
	hidden_states = self.dense_4h_to_h(hidden_states)
	return hidden_states


	class GPTNeoXLayer(nn.Module):
	def __init__(self, config):
	super().__init__()
	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.attention = GPTNeoXAttention(config)
	self.mlp = GPTNeoXMLP(config)

	def forward(
	self,
	hidden_states,
	attention_mask=None,
	head_mask=None,
	use_cache=False,
	layer_past=None,
	output_attentions=False,
	):
	residual = hidden_states
	ln_out = self.input_layernorm(hidden_states)
	attention_layer_outputs = self.attention(
	ln_out,
	attention_mask=attention_mask,
	layer_past=layer_past,
	head_mask=head_mask,
	use_cache=use_cache,
	output_attentions=output_attentions,
	)
	attn_output = attention_layer_outputs[0] # output_attn: a, present, (attentions)
	outputs = attention_layer_outputs[1:]

	mlp_output = self.mlp(self.post_attention_layernorm(hidden_states))
	hidden_states = mlp_output + attn_output + residual

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

	return outputs # hidden_states, present, (attentions)


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

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

	# 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=None,
	attention_mask=None,
	head_mask=None,
	inputs_embeds=None,
	past_key_values=None,
	use_cache=None,
	output_attentions=None,
	output_hidden_states=None,
	return_dict=None,
	):
	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:
	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_key_values = tuple([None] * self.config.num_hidden_layers)

	# 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 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) * -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
	# 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 = inputs_embeds

	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,)
	outputs = layer(
	hidden_states,
	attention_mask=attention_mask,
	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 GPTNeoXForCausalLM(GPTNeoXPreTrainedModel, GenerationMixin):

	_keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]

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

	self.gpt_neox = GPTNeoXModel(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=None,
	attention_mask=None,
	inputs_embeds=None,
	head_mask=None,
	past_key_values=None,
	labels=None,
	use_cache=None,
	output_attentions=None,
	output_hidden_states=None,
	return_dict=None,
	):
	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 GPTNeoXTokenizer, GPTNeoXForCausalLM, GPTNeoXConfig
	>>> import torch

	>>> tokenizer = GPTNeoXTokenizer.from_pretrained("gpt-neox-20b")
	>>> config = GPTNeoXConfig.from_pretrained("gpt-neox-20b")
	>>> config.is_decoder = True
	>>> model = GPTNeoXForCausalLM.from_pretrained("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,
	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:
	# 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=None, attention_mask=None, **model_kwargs):
	input_shape = input_ids.shape

	# 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)

	# cut decoder_input_ids if past is used
	if past is not None:
	input_ids = input_ids[:, -1:]

	return {"input_ids": input_ids, "attention_mask": attention_mask, "past_key_values": past}

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