Package upstream model with reproducible local MMEB evaluation

bb2b5cc verified 12 days ago

20.4 kB

# Copyright 2022 Salesforce authors, The EleutherAI, and HuggingFace Teams. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""PyTorch CodeGen model."""

import math

import torch
from torch import nn

from ... import initialization as init
from ...activations import ACT2FN
from ...cache_utils import Cache, DynamicCache
from ...generation import GenerationMixin
from ...masking_utils import create_causal_mask
from ...modeling_layers import GradientCheckpointingLayer
from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
from ...modeling_utils import PreTrainedModel
from ...utils import (
    auto_docstring,
    logging,
)
from .configuration_codegen import CodeGenConfig


logger = logging.get_logger(__name__)


# Copied from transformers.models.gptj.modeling_gptj.create_sinusoidal_positions
def create_sinusoidal_positions(num_pos: int, dim: int) -> torch.Tensor:
    inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2, dtype=torch.int64) / dim))
    sinusoid_inp = torch.einsum("i , j -> i j", torch.arange(num_pos, dtype=torch.int64).float(), inv_freq).float()
    return torch.cat((torch.sin(sinusoid_inp), torch.cos(sinusoid_inp)), dim=1)


# Copied from transformers.models.gptj.modeling_gptj.rotate_every_two
def rotate_every_two(x: torch.Tensor) -> torch.Tensor:
    x1 = x[:, :, :, ::2]
    x2 = x[:, :, :, 1::2]
    x = torch.stack((-x2, x1), dim=-1)
    return x.flatten(-2)  # in einsum notation: rearrange(x, '... d j -> ... (d j)')


# Copied from transformers.models.gptj.modeling_gptj.apply_rotary_pos_emb
def apply_rotary_pos_emb(tensor: torch.Tensor, sin: torch.Tensor, cos: torch.Tensor) -> torch.Tensor:
    sin = torch.repeat_interleave(sin[:, :, None, :], 2, 3)
    cos = torch.repeat_interleave(cos[:, :, None, :], 2, 3)
    return (tensor * cos) + (rotate_every_two(tensor) * sin)


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

        self.max_positions = config.max_position_embeddings
        self.attn_dropout = nn.Dropout(config.attn_pdrop)
        self.resid_dropout = nn.Dropout(config.resid_pdrop)
        self.layer_idx = layer_idx
        if layer_idx is None:
            logger.warning_once(
                f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
                "when creating this class."
            )

        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 = math.sqrt(self.head_dim)
        self.qkv_proj = nn.Linear(self.embed_dim, self.embed_dim * 3, bias=False)

        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)
        self.rotary_dim = config.rotary_dim
        self.pos_embd_dim = self.rotary_dim or self.embed_dim
        self.register_buffer(
            "embed_positions", create_sinusoidal_positions(self.max_positions, self.pos_embd_dim), persistent=False
        )

    def _split_heads(self, x, n_head, dim_head, mp_num):
        reshaped = x.reshape(x.shape[:-1] + (n_head // mp_num, dim_head))
        reshaped = reshaped.reshape(x.shape[:-2] + (-1,) + reshaped.shape[-1:])
        return reshaped

    def _merge_heads(self, tensor, num_attention_heads, attn_head_size):
        """
        Merges attn_head_size dim and num_attn_heads dim into n_ctx
        """
        if len(tensor.shape) == 5:
            tensor = tensor.permute(0, 1, 3, 2, 4).contiguous()
        elif len(tensor.shape) == 4:
            tensor = tensor.permute(0, 2, 1, 3).contiguous()
        else:
            raise ValueError(f"Input tensor rank should be one of [4, 5], but is: {len(tensor.shape)}")
        new_shape = tensor.size()[:-2] + (num_attention_heads * attn_head_size,)
        return tensor.view(new_shape)

    def _attn(
        self,
        query,
        key,
        value,
        attention_mask=None,
    ):
        # 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))

        if attention_mask is not None:
            attn_weights = attn_weights + attention_mask

        attn_weights = attn_weights / self.scale_attn
        attn_weights = nn.Softmax(dim=-1)(attn_weights)
        attn_weights = attn_weights.to(value.dtype)
        attn_weights = self.attn_dropout(attn_weights)

        attn_output = torch.matmul(attn_weights, value)

        return attn_output, attn_weights

    def forward(
        self,
        hidden_states: torch.FloatTensor | None,
        layer_past: Cache | None = None,
        attention_mask: torch.FloatTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        use_cache: bool | None = False,
        output_attentions: bool | None = False,
    ) -> (
        tuple[torch.Tensor, tuple[torch.Tensor]]
        | tuple[torch.Tensor, tuple[torch.Tensor], tuple[torch.Tensor, ...]]
        | None
    ):
        qkv = self.qkv_proj(hidden_states)
        # TODO(enijkamp): factor out number of logical TPU-v4 cores or make forward pass agnostic
        mp_num = 4
        qkv_split = qkv.reshape(qkv.shape[:-1] + (mp_num, -1))

        local_dim = self.head_dim * self.num_attention_heads // mp_num
        query, value, key = torch.split(qkv_split, local_dim, dim=-1)
        query = self._split_heads(query, self.num_attention_heads, self.head_dim, mp_num=mp_num)
        key = self._split_heads(key, self.num_attention_heads, self.head_dim, mp_num=mp_num)

        value = self._split_heads(value, self.num_attention_heads, self.head_dim, mp_num=mp_num)
        value = value.permute(0, 2, 1, 3)

        embed_positions = self.embed_positions
        if embed_positions.device != position_ids.device:
            embed_positions = embed_positions.to(position_ids.device)
            self.embed_positions = embed_positions

        sincos = embed_positions[position_ids]
        sin, cos = torch.split(sincos, sincos.shape[-1] // 2, dim=-1)

        if self.rotary_dim is not None:
            k_rot = key[:, :, :, : self.rotary_dim]
            k_pass = key[:, :, :, self.rotary_dim :]

            q_rot = query[:, :, :, : self.rotary_dim]
            q_pass = query[:, :, :, self.rotary_dim :]

            k_rot = apply_rotary_pos_emb(k_rot, sin, cos)
            q_rot = apply_rotary_pos_emb(q_rot, sin, cos)

            key = torch.cat([k_rot, k_pass], dim=-1)
            query = torch.cat([q_rot, q_pass], dim=-1)
        else:
            key = apply_rotary_pos_emb(key, sin, cos)
            query = apply_rotary_pos_emb(query, sin, cos)

        key = key.permute(0, 2, 1, 3)
        query = query.permute(0, 2, 1, 3)

        # Note that this cast is quite ugly, but is not implemented before ROPE as k_rot in the original codebase is always in fp32.
        # Reference: https://github.com/salesforce/CodeGen/blob/f210c3bb1216c975ad858cd4132c0fdeabf4bfc2/codegen1/jaxformer/hf/codegen/modeling_codegen.py#L38
        if layer_past is not None:
            key, value = layer_past.update(key.to(hidden_states.dtype), value, self.layer_idx)

        # compute self-attention: V x Softmax(QK^T)
        attn_output, attn_weights = self._attn(query, key, value, attention_mask)

        attn_output = self._merge_heads(attn_output, self.num_attention_heads, self.head_dim)
        attn_output = self.out_proj(attn_output)
        attn_output = self.resid_dropout(attn_output)
        return attn_output, attn_weights


# Copied from transformers.models.gptj.modeling_gptj.GPTJMLP with GPTJ->CodeGen
class CodeGenMLP(nn.Module):
    def __init__(self, intermediate_size, config):  # in MLP: intermediate_size= 4 * embed_dim
        super().__init__()
        embed_dim = config.n_embd

        self.fc_in = nn.Linear(embed_dim, intermediate_size)
        self.fc_out = nn.Linear(intermediate_size, embed_dim)

        self.act = ACT2FN[config.activation_function]
        self.dropout = nn.Dropout(config.resid_pdrop)

    def forward(self, hidden_states: torch.FloatTensor | None) -> torch.FloatTensor:
        hidden_states = self.fc_in(hidden_states)
        hidden_states = self.act(hidden_states)
        hidden_states = self.fc_out(hidden_states)
        hidden_states = self.dropout(hidden_states)
        return hidden_states


# Copied from transformers.models.gptj.modeling_gptj.GPTJBlock with GPTJ->CodeGen
class CodeGenBlock(GradientCheckpointingLayer):
    # Ignore copy
    def __init__(self, config, layer_idx=None):
        super().__init__()
        inner_dim = config.n_inner if config.n_inner is not None else 4 * config.n_embd
        self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
        self.attn = CodeGenAttention(config, layer_idx)
        self.mlp = CodeGenMLP(inner_dim, config)

    def forward(
        self,
        hidden_states: torch.FloatTensor | None,
        layer_past: Cache | None = None,
        attention_mask: torch.FloatTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        use_cache: bool | None = False,
        output_attentions: bool | None = False,
        **kwargs,
    ) -> tuple[torch.Tensor] | tuple[torch.Tensor, tuple[torch.FloatTensor, ...]] | None:
        residual = hidden_states
        hidden_states = self.ln_1(hidden_states)
        attn_outputs, attn_weights = self.attn(
            hidden_states=hidden_states,
            layer_past=layer_past,
            attention_mask=attention_mask,
            position_ids=position_ids,
            use_cache=use_cache,
            output_attentions=output_attentions,
        )
        feed_forward_hidden_states = self.mlp(hidden_states)
        hidden_states = attn_outputs + feed_forward_hidden_states + residual

        return hidden_states, attn_weights


@auto_docstring
class CodeGenPreTrainedModel(PreTrainedModel):
    config: CodeGenConfig
    base_model_prefix = "transformer"
    supports_gradient_checkpointing = True
    _no_split_modules = ["CodeGenBlock"]
    _skip_keys_device_placement = "past_key_values"
    _can_compile_fullgraph = True

    def _init_weights(self, module):
        super()._init_weights(module)
        if isinstance(module, CodeGenAttention):
            init.copy_(module.embed_positions, create_sinusoidal_positions(module.max_positions, module.pos_embd_dim))


@auto_docstring
class CodeGenModel(CodeGenPreTrainedModel):
    def __init__(self, config):
        super().__init__(config)

        self.embed_dim = config.n_embd
        self.vocab_size = config.vocab_size
        self.wte = nn.Embedding(config.vocab_size, self.embed_dim)
        self.drop = nn.Dropout(config.embd_pdrop)
        self.h = nn.ModuleList([CodeGenBlock(config, layer_idx=i) for i in range(config.n_layer)])
        self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
        self.rotary_dim = min(config.rotary_dim, config.n_ctx // config.num_attention_heads)

        self.gradient_checkpointing = False

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

    def get_input_embeddings(self):
        return self.wte

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

    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        attention_mask: torch.FloatTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        use_cache: bool | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs,  # NOOP kwargs, for now
    ) -> tuple | BaseModelOutputWithPast:
        r"""
        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_dim)`, *optional*):
            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
            is useful if you want more control over how to convert *input_ids* indices into associated vectors than the
            model's internal embedding lookup matrix.
        """
        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.return_dict

        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

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

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

        if use_cache and past_key_values is None:
            past_key_values = DynamicCache(config=self.config)

        seq_length = inputs_embeds.shape[1]
        if position_ids is None:
            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
            position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen_tokens
            position_ids = position_ids.unsqueeze(0)

        causal_mask = create_causal_mask(
            config=self.config,
            inputs_embeds=inputs_embeds,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            position_ids=position_ids,
        )

        hidden_states = inputs_embeds

        if token_type_ids is not None:
            token_type_ids = token_type_ids.view(-1, seq_length)
            token_type_embeds = self.wte(token_type_ids)
            hidden_states = hidden_states + token_type_embeds

        hidden_states = self.drop(hidden_states)
        output_shape = (-1, seq_length, hidden_states.size(-1))

        all_self_attentions = () if output_attentions else None
        all_hidden_states = () if output_hidden_states else None
        for i, block in enumerate(self.h):
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            outputs = block(
                hidden_states,
                layer_past=past_key_values,
                attention_mask=causal_mask,
                position_ids=position_ids,
                use_cache=use_cache,
                output_attentions=output_attentions,
            )

            hidden_states = outputs[0]
            if output_attentions:
                all_self_attentions = all_self_attentions + (outputs[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, past_key_values, all_hidden_states, all_self_attentions] if v is not None
            )

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


@auto_docstring(
    custom_intro="""
    The CodeGen Model transformer with a language modeling head on top.
    """
)
class CodeGenForCausalLM(CodeGenPreTrainedModel, GenerationMixin):
    _tied_weights_keys = {"lm_head.weight": "transformer.wte.weight"}

    def __init__(self, config):
        super().__init__(config)
        self.transformer = CodeGenModel(config)
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size)

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

    @auto_docstring
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        past_key_values: Cache | None = None,
        attention_mask: torch.FloatTensor | None = None,
        token_type_ids: torch.LongTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        labels: torch.LongTensor | None = None,
        use_cache: bool | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        return_dict: bool | None = None,
        logits_to_keep: int | torch.Tensor = 0,
        **kwargs,
    ) -> tuple | CausalLMOutputWithPast:
        r"""
        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_dim)`, *optional*):
            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
            is useful if you want more control over how to convert *input_ids* indices into associated vectors than the
            model's internal embedding lookup matrix.
        labels (`torch.LongTensor` of shape `(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.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,
            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[0]
        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        logits = self.lm_head(hidden_states[:, slice_indices, :])

        loss = None
        if labels is not None:
            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)

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

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=transformer_outputs.past_key_values,
            hidden_states=transformer_outputs.hidden_states,
            attentions=transformer_outputs.attentions,
        )


__all__ = ["CodeGenForCausalLM", "CodeGenModel", "CodeGenPreTrainedModel"]

	# Copyright 2022 Salesforce authors, The EleutherAI, and HuggingFace Teams. All rights reserved.
	#
	# Licensed under the Apache License, Version 2.0 (the "License");
	# you may not use this file except in compliance with the License.
	# You may obtain a copy of the License at
	#
	# http://www.apache.org/licenses/LICENSE-2.0
	#
	# Unless required by applicable law or agreed to in writing, software
	# distributed under the License is distributed on an "AS IS" BASIS,
	# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
	# See the License for the specific language governing permissions and
	# limitations under the License.
	"""PyTorch CodeGen model."""

	import math

	import torch
	from torch import nn

	from ... import initialization as init
	from ...activations import ACT2FN
	from ...cache_utils import Cache, DynamicCache
	from ...generation import GenerationMixin
	from ...masking_utils import create_causal_mask
	from ...modeling_layers import GradientCheckpointingLayer
	from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
	from ...modeling_utils import PreTrainedModel
	from ...utils import (
	auto_docstring,
	logging,
	)
	from .configuration_codegen import CodeGenConfig


	logger = logging.get_logger(__name__)


	# Copied from transformers.models.gptj.modeling_gptj.create_sinusoidal_positions
	def create_sinusoidal_positions(num_pos: int, dim: int) -> torch.Tensor:
	inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2, dtype=torch.int64) / dim))
	sinusoid_inp = torch.einsum("i , j -> i j", torch.arange(num_pos, dtype=torch.int64).float(), inv_freq).float()
	return torch.cat((torch.sin(sinusoid_inp), torch.cos(sinusoid_inp)), dim=1)


	# Copied from transformers.models.gptj.modeling_gptj.rotate_every_two
	def rotate_every_two(x: torch.Tensor) -> torch.Tensor:
	x1 = x[:, :, :, ::2]
	x2 = x[:, :, :, 1::2]
	x = torch.stack((-x2, x1), dim=-1)
	return x.flatten(-2) # in einsum notation: rearrange(x, '... d j -> ... (d j)')


	# Copied from transformers.models.gptj.modeling_gptj.apply_rotary_pos_emb
	def apply_rotary_pos_emb(tensor: torch.Tensor, sin: torch.Tensor, cos: torch.Tensor) -> torch.Tensor:
	sin = torch.repeat_interleave(sin[:, :, None, :], 2, 3)
	cos = torch.repeat_interleave(cos[:, :, None, :], 2, 3)
	return (tensor * cos) + (rotate_every_two(tensor) * sin)


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

	self.max_positions = config.max_position_embeddings
	self.attn_dropout = nn.Dropout(config.attn_pdrop)
	self.resid_dropout = nn.Dropout(config.resid_pdrop)
	self.layer_idx = layer_idx
	if layer_idx is None:
	logger.warning_once(
	f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
	"lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
	"when creating this class."
	)

	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 = math.sqrt(self.head_dim)
	self.qkv_proj = nn.Linear(self.embed_dim, self.embed_dim * 3, bias=False)

	self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)
	self.rotary_dim = config.rotary_dim
	self.pos_embd_dim = self.rotary_dim or self.embed_dim
	self.register_buffer(
	"embed_positions", create_sinusoidal_positions(self.max_positions, self.pos_embd_dim), persistent=False
	)

	def _split_heads(self, x, n_head, dim_head, mp_num):
	reshaped = x.reshape(x.shape[:-1] + (n_head // mp_num, dim_head))
	reshaped = reshaped.reshape(x.shape[:-2] + (-1,) + reshaped.shape[-1:])
	return reshaped

	def _merge_heads(self, tensor, num_attention_heads, attn_head_size):
	"""
	Merges attn_head_size dim and num_attn_heads dim into n_ctx
	"""
	if len(tensor.shape) == 5:
	tensor = tensor.permute(0, 1, 3, 2, 4).contiguous()
	elif len(tensor.shape) == 4:
	tensor = tensor.permute(0, 2, 1, 3).contiguous()
	else:
	raise ValueError(f"Input tensor rank should be one of [4, 5], but is: {len(tensor.shape)}")
	new_shape = tensor.size()[:-2] + (num_attention_heads * attn_head_size,)
	return tensor.view(new_shape)

	def _attn(
	self,
	query,
	key,
	value,
	attention_mask=None,
	):
	# 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))

	if attention_mask is not None:
	attn_weights = attn_weights + attention_mask

	attn_weights = attn_weights / self.scale_attn
	attn_weights = nn.Softmax(dim=-1)(attn_weights)
	attn_weights = attn_weights.to(value.dtype)
	attn_weights = self.attn_dropout(attn_weights)

	attn_output = torch.matmul(attn_weights, value)

	return attn_output, attn_weights

	def forward(
	self,
	hidden_states: torch.FloatTensor \| None,
	layer_past: Cache \| None = None,
	attention_mask: torch.FloatTensor \| None = None,
	position_ids: torch.LongTensor \| None = None,
	use_cache: bool \| None = False,
	output_attentions: bool \| None = False,
	) -> (
	tuple[torch.Tensor, tuple[torch.Tensor]]
	\| tuple[torch.Tensor, tuple[torch.Tensor], tuple[torch.Tensor, ...]]
	\| None
	):
	qkv = self.qkv_proj(hidden_states)
	# TODO(enijkamp): factor out number of logical TPU-v4 cores or make forward pass agnostic
	mp_num = 4
	qkv_split = qkv.reshape(qkv.shape[:-1] + (mp_num, -1))

	local_dim = self.head_dim * self.num_attention_heads // mp_num
	query, value, key = torch.split(qkv_split, local_dim, dim=-1)
	query = self._split_heads(query, self.num_attention_heads, self.head_dim, mp_num=mp_num)
	key = self._split_heads(key, self.num_attention_heads, self.head_dim, mp_num=mp_num)

	value = self._split_heads(value, self.num_attention_heads, self.head_dim, mp_num=mp_num)
	value = value.permute(0, 2, 1, 3)

	embed_positions = self.embed_positions
	if embed_positions.device != position_ids.device:
	embed_positions = embed_positions.to(position_ids.device)
	self.embed_positions = embed_positions

	sincos = embed_positions[position_ids]
	sin, cos = torch.split(sincos, sincos.shape[-1] // 2, dim=-1)

	if self.rotary_dim is not None:
	k_rot = key[:, :, :, : self.rotary_dim]
	k_pass = key[:, :, :, self.rotary_dim :]

	q_rot = query[:, :, :, : self.rotary_dim]
	q_pass = query[:, :, :, self.rotary_dim :]

	k_rot = apply_rotary_pos_emb(k_rot, sin, cos)
	q_rot = apply_rotary_pos_emb(q_rot, sin, cos)

	key = torch.cat([k_rot, k_pass], dim=-1)
	query = torch.cat([q_rot, q_pass], dim=-1)
	else:
	key = apply_rotary_pos_emb(key, sin, cos)
	query = apply_rotary_pos_emb(query, sin, cos)

	key = key.permute(0, 2, 1, 3)
	query = query.permute(0, 2, 1, 3)

	# Note that this cast is quite ugly, but is not implemented before ROPE as k_rot in the original codebase is always in fp32.
	# Reference: https://github.com/salesforce/CodeGen/blob/f210c3bb1216c975ad858cd4132c0fdeabf4bfc2/codegen1/jaxformer/hf/codegen/modeling_codegen.py#L38
	if layer_past is not None:
	key, value = layer_past.update(key.to(hidden_states.dtype), value, self.layer_idx)

	# compute self-attention: V x Softmax(QK^T)
	attn_output, attn_weights = self._attn(query, key, value, attention_mask)

	attn_output = self._merge_heads(attn_output, self.num_attention_heads, self.head_dim)
	attn_output = self.out_proj(attn_output)
	attn_output = self.resid_dropout(attn_output)
	return attn_output, attn_weights


	# Copied from transformers.models.gptj.modeling_gptj.GPTJMLP with GPTJ->CodeGen
	class CodeGenMLP(nn.Module):
	def __init__(self, intermediate_size, config): # in MLP: intermediate_size= 4 * embed_dim
	super().__init__()
	embed_dim = config.n_embd

	self.fc_in = nn.Linear(embed_dim, intermediate_size)
	self.fc_out = nn.Linear(intermediate_size, embed_dim)

	self.act = ACT2FN[config.activation_function]
	self.dropout = nn.Dropout(config.resid_pdrop)

	def forward(self, hidden_states: torch.FloatTensor \| None) -> torch.FloatTensor:
	hidden_states = self.fc_in(hidden_states)
	hidden_states = self.act(hidden_states)
	hidden_states = self.fc_out(hidden_states)
	hidden_states = self.dropout(hidden_states)
	return hidden_states


	# Copied from transformers.models.gptj.modeling_gptj.GPTJBlock with GPTJ->CodeGen
	class CodeGenBlock(GradientCheckpointingLayer):
	# Ignore copy
	def __init__(self, config, layer_idx=None):
	super().__init__()
	inner_dim = config.n_inner if config.n_inner is not None else 4 * config.n_embd
	self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
	self.attn = CodeGenAttention(config, layer_idx)
	self.mlp = CodeGenMLP(inner_dim, config)

	def forward(
	self,
	hidden_states: torch.FloatTensor \| None,
	layer_past: Cache \| None = None,
	attention_mask: torch.FloatTensor \| None = None,
	position_ids: torch.LongTensor \| None = None,
	use_cache: bool \| None = False,
	output_attentions: bool \| None = False,
	**kwargs,
	) -> tuple[torch.Tensor] \| tuple[torch.Tensor, tuple[torch.FloatTensor, ...]] \| None:
	residual = hidden_states
	hidden_states = self.ln_1(hidden_states)
	attn_outputs, attn_weights = self.attn(
	hidden_states=hidden_states,
	layer_past=layer_past,
	attention_mask=attention_mask,
	position_ids=position_ids,
	use_cache=use_cache,
	output_attentions=output_attentions,
	)
	feed_forward_hidden_states = self.mlp(hidden_states)
	hidden_states = attn_outputs + feed_forward_hidden_states + residual

	return hidden_states, attn_weights


	@auto_docstring
	class CodeGenPreTrainedModel(PreTrainedModel):
	config: CodeGenConfig
	base_model_prefix = "transformer"
	supports_gradient_checkpointing = True
	_no_split_modules = ["CodeGenBlock"]
	_skip_keys_device_placement = "past_key_values"
	_can_compile_fullgraph = True

	def _init_weights(self, module):
	super()._init_weights(module)
	if isinstance(module, CodeGenAttention):
	init.copy_(module.embed_positions, create_sinusoidal_positions(module.max_positions, module.pos_embd_dim))


	@auto_docstring
	class CodeGenModel(CodeGenPreTrainedModel):
	def __init__(self, config):
	super().__init__(config)

	self.embed_dim = config.n_embd
	self.vocab_size = config.vocab_size
	self.wte = nn.Embedding(config.vocab_size, self.embed_dim)
	self.drop = nn.Dropout(config.embd_pdrop)
	self.h = nn.ModuleList([CodeGenBlock(config, layer_idx=i) for i in range(config.n_layer)])
	self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
	self.rotary_dim = min(config.rotary_dim, config.n_ctx // config.num_attention_heads)

	self.gradient_checkpointing = False

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

	def get_input_embeddings(self):
	return self.wte

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

	@auto_docstring
	def forward(
	self,
	input_ids: torch.LongTensor \| None = None,
	past_key_values: Cache \| None = None,
	attention_mask: torch.FloatTensor \| None = None,
	token_type_ids: torch.LongTensor \| None = None,
	position_ids: torch.LongTensor \| None = None,
	inputs_embeds: torch.FloatTensor \| None = None,
	use_cache: bool \| None = None,
	output_attentions: bool \| None = None,
	output_hidden_states: bool \| None = None,
	return_dict: bool \| None = None,
	**kwargs, # NOOP kwargs, for now
	) -> tuple \| BaseModelOutputWithPast:
	r"""
	inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_dim)`, optional):
	Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
	is useful if you want more control over how to convert input_ids indices into associated vectors than the
	model's internal embedding lookup matrix.
	"""
	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.return_dict

	if (input_ids is None) ^ (inputs_embeds is not None):
	raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

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

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

	if use_cache and past_key_values is None:
	past_key_values = DynamicCache(config=self.config)

	seq_length = inputs_embeds.shape[1]
	if position_ids is None:
	past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
	position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen_tokens
	position_ids = position_ids.unsqueeze(0)

	causal_mask = create_causal_mask(
	config=self.config,
	inputs_embeds=inputs_embeds,
	attention_mask=attention_mask,
	past_key_values=past_key_values,
	position_ids=position_ids,
	)

	hidden_states = inputs_embeds

	if token_type_ids is not None:
	token_type_ids = token_type_ids.view(-1, seq_length)
	token_type_embeds = self.wte(token_type_ids)
	hidden_states = hidden_states + token_type_embeds

	hidden_states = self.drop(hidden_states)
	output_shape = (-1, seq_length, hidden_states.size(-1))

	all_self_attentions = () if output_attentions else None
	all_hidden_states = () if output_hidden_states else None
	for i, block in enumerate(self.h):
	if output_hidden_states:
	all_hidden_states = all_hidden_states + (hidden_states,)

	outputs = block(
	hidden_states,
	layer_past=past_key_values,
	attention_mask=causal_mask,
	position_ids=position_ids,
	use_cache=use_cache,
	output_attentions=output_attentions,
	)

	hidden_states = outputs[0]
	if output_attentions:
	all_self_attentions = all_self_attentions + (outputs[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, past_key_values, all_hidden_states, all_self_attentions] if v is not None
	)

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


	@auto_docstring(
	custom_intro="""
	The CodeGen Model transformer with a language modeling head on top.
	"""
	)
	class CodeGenForCausalLM(CodeGenPreTrainedModel, GenerationMixin):
	_tied_weights_keys = {"lm_head.weight": "transformer.wte.weight"}

	def __init__(self, config):
	super().__init__(config)
	self.transformer = CodeGenModel(config)
	self.lm_head = nn.Linear(config.n_embd, config.vocab_size)

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

	@auto_docstring
	def forward(
	self,
	input_ids: torch.LongTensor \| None = None,
	past_key_values: Cache \| None = None,
	attention_mask: torch.FloatTensor \| None = None,
	token_type_ids: torch.LongTensor \| None = None,
	position_ids: torch.LongTensor \| None = None,
	inputs_embeds: torch.FloatTensor \| None = None,
	labels: torch.LongTensor \| None = None,
	use_cache: bool \| None = None,
	output_attentions: bool \| None = None,
	output_hidden_states: bool \| None = None,
	return_dict: bool \| None = None,
	logits_to_keep: int \| torch.Tensor = 0,
	**kwargs,
	) -> tuple \| CausalLMOutputWithPast:
	r"""
	inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_dim)`, optional):
	Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
	is useful if you want more control over how to convert input_ids indices into associated vectors than the
	model's internal embedding lookup matrix.
	labels (`torch.LongTensor` of shape `(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.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,
	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[0]
	# Only compute necessary logits, and do not upcast them to float if we are not computing the loss
	slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
	logits = self.lm_head(hidden_states[:, slice_indices, :])

	loss = None
	if labels is not None:
	loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)

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

	return CausalLMOutputWithPast(
	loss=loss,
	logits=logits,
	past_key_values=transformer_outputs.past_key_values,
	hidden_states=transformer_outputs.hidden_states,
	attentions=transformer_outputs.attentions,
	)


	__all__ = ["CodeGenForCausalLM", "CodeGenModel", "CodeGenPreTrainedModel"]