ColossalAI/colossalai/shardformer/modeling/gptj.py

from typing import Dict, List, Optional, Tuple, Union

import torch
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
from transformers.modeling_outputs import (
    BaseModelOutputWithPast,
    CausalLMOutputWithPast,
    QuestionAnsweringModelOutput,
    SequenceClassifierOutputWithPast,
)
from transformers.models.gptj.modeling_gptj import (
    GPTJForCausalLM,
    GPTJForQuestionAnswering,
    GPTJForSequenceClassification,
    GPTJModel,
    apply_rotary_pos_emb,
    get_embed_positions,
)
from transformers.utils import is_torch_fx_proxy, logging

from colossalai.pipeline.stage_manager import PipelineStageManager
from colossalai.shardformer.layer import ColoAttention
from colossalai.shardformer.layer._operation import gather_forward_split_backward, split_forward_gather_backward
from colossalai.shardformer.shard import ShardConfig

logger = logging.get_logger(__name__)


def _get_attention_mask(
    self: GPTJModel,
    shard_config: ShardConfig,
    hidden_states: torch.Tensor,
    past_key_values: Optional[Tuple[Tuple[torch.Tensor]]],
    attention_mask: Optional[torch.FloatTensor],
    use_flash_attention_2: bool = False,
) -> Optional[Union[torch.Tensor, dict]]:
    batch_size, seq_len = hidden_states.shape[:2]
    past_key_values_length = 0
    if past_key_values is not None and past_key_values[0] is not None:
        past_key_values_length = past_key_values[0][0].shape[2]
    if shard_config.enable_flash_attention:
        if attention_mask is not None:
            attention_mask = attention_mask.view(batch_size, -1)
        attention_mask = ColoAttention.prepare_attn_kwargs(
            (batch_size, 1, seq_len, seq_len + past_key_values_length),
            hidden_states.dtype,
            hidden_states.device,
            attention_mask,
            is_causal=True,
        )
    elif use_flash_attention_2 and attention_mask is not None:
        if batch_size <= 0:
            raise ValueError("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
    return attention_mask


class GPTJPipelineForwards:
    """
    This class serves as a micro library for forward function substitution of GPTJ models
    under pipeline setting.
    """

    @staticmethod
    def gptj_model_forward(
        self: GPTJModel,
        input_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        token_type_ids: Optional[torch.LongTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        stage_manager: Optional[PipelineStageManager] = None,
        hidden_states: Optional[torch.FloatTensor] = None,
        stage_index: Optional[List[int]] = None,
        shard_config: ShardConfig = None,
    ) -> Union[Dict, Tuple, BaseModelOutputWithPast]:
        # This function is modified on the basis of transformers.models.gptj.modeling_gptj.GPTJModel.forward.
        # Please refer to original code of transformers for more details.
        # GPTJ has no cross attention in comparison to GPT2

        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        logger = logging.get_logger(__name__)

        # Preprocess passed in arguments
        # TODO(baizhou): left the recording kv-value tensors as () or None type, this feature may be added in the future.
        if past_key_values:
            logger.warning_once("Non-empty past_key_values is not supported for pipeline models at the moment.")
            past_key_values = None
        if output_attentions:
            logger.warning_once("output_attentions=True is not supported for pipeline models at the moment.")
            output_attentions = False
        if output_hidden_states:
            logger.warning_once("output_hidden_states=True is not supported for pipeline models at the moment.")
            output_hidden_states = False
        if use_cache:
            logger.warning_once("use_cache=True is not supported for pipeline models at the moment.")
            use_cache = False

        if stage_manager.is_first_stage():
            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:
                batch_size, seq_length = input_ids.shape
                input_shape = input_ids.size()
                input_ids = input_ids.view(-1, seq_length)

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

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

            if token_type_ids is not None:
                token_type_ids = token_type_ids.view(-1, seq_length)
        else:
            if hidden_states is None:
                raise ValueError("hidden_states shouldn't be None for stages other than the first stage.")
            input_shape = hidden_states.size()[:-1]
            batch_size, seq_length = input_shape[0], input_shape[1]
            device = hidden_states.device

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

        # position id to be assigned not just for the first stage for attn input
        if position_ids is None:
            position_ids = torch.arange(0, seq_length, dtype=torch.long, device=device)
            position_ids = position_ids.unsqueeze(0)
        if stage_manager.is_first_stage():
            if inputs_embeds is None:
                inputs_embeds = self.wte(input_ids)
            hidden_states = inputs_embeds
            if token_type_ids is not None:
                token_type_embeds = self.wte(token_type_ids)
                hidden_states = hidden_states + token_type_embeds
            hidden_states = self.drop(hidden_states)

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

        attention_mask = _get_attention_mask(
            self, shard_config, hidden_states, past_key_values, attention_mask, self._use_flash_attention_2
        )

        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

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

        # split the input tensor along sequence dimension
        # [batch_size, seq_len, hidden_size] -> [batch_size, seq_len/TP_size, hidden_size]
        if shard_config.enable_sequence_parallelism:
            hidden_states = split_forward_gather_backward(
                hidden_states,
                dim=1,
                process_group=shard_config.tensor_parallel_process_group,
            )

        # Going through held blocks.
        start_idx, end_idx = stage_index[0], stage_index[1]
        for i in range(start_idx, end_idx):
            block = self.h[i]
            torch.cuda.set_device(hidden_states.device)

            # Ensure that attention_mask is always on the same device as hidden_states
            if attention_mask is not None:
                attention_mask = attention_mask.to(hidden_states.device)
            if isinstance(head_mask, torch.Tensor):
                head_mask = head_mask.to(hidden_states.device)
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            if self.gradient_checkpointing and self.training:
                outputs = self._gradient_checkpointing_func(
                    block.__call__,
                    hidden_states,
                    None,
                    attention_mask,
                    position_ids,
                    head_mask[i],
                    use_cache,
                    output_attentions,
                )
            else:
                outputs = block(
                    hidden_states=hidden_states,
                    layer_past=None,
                    attention_mask=attention_mask,
                    position_ids=position_ids,
                    head_mask=head_mask[i],
                    use_cache=use_cache,
                    output_attentions=output_attentions,
                )

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

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

        # When sequence parallelism done, gather the output tensor in forward and split it in backward
        if shard_config.enable_sequence_parallelism:
            hidden_states = gather_forward_split_backward(
                hidden_states,
                dim=1,
                process_group=shard_config.tensor_parallel_process_group,
            )

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

            return BaseModelOutputWithPast(
                last_hidden_state=hidden_states,
                past_key_values=presents,
                hidden_states=all_hidden_states,
                attentions=all_self_attentions,
            )
        else:
            # always return dict for intermediate stage
            return {"hidden_states": hidden_states}

    @staticmethod
    def gptj_causallm_model_forward(
        self: GPTJForCausalLM,
        input_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        token_type_ids: Optional[torch.LongTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        stage_manager: Optional[PipelineStageManager] = None,
        hidden_states: Optional[torch.FloatTensor] = None,
        stage_index: Optional[List[int]] = None,
        shard_config: ShardConfig = None,
    ) -> Union[Dict, Tuple, CausalLMOutputWithPast]:
        r"""
        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]`

        # This function is modified on the basis of transformers.models.gptj.modeling_gptj.GPTJForCausalLM.forward.
        # Please refer to original code of transformers for more details.
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        transformer_outputs = GPTJPipelineForwards.gptj_model_forward(
            self.transformer,
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            stage_manager=stage_manager,
            hidden_states=hidden_states,
            stage_index=stage_index,
            shard_config=shard_config,
        )

        # If not at the last stage, return hidden_states as in GPTJModel
        if not stage_manager.is_last_stage():
            return {"hidden_states": transformer_outputs["hidden_states"]}

        hidden_states = transformer_outputs[0]
        lm_logits = self.lm_head(hidden_states)

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

            loss = loss.to(hidden_states.dtype)

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

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

    @staticmethod
    def gptj_for_sequence_classification_forward(
        self: GPTJForSequenceClassification,
        input_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        token_type_ids: Optional[torch.LongTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        stage_manager: Optional[PipelineStageManager] = None,
        hidden_states: Optional[torch.FloatTensor] = None,
        stage_index: Optional[List[int]] = None,
        shard_config: ShardConfig = None,
    ) -> Union[Dict, Tuple, SequenceClassifierOutputWithPast]:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).

        # This function is modified on the basis of transformers.models.gptj.modeling_gptj.GPTJForSequenceClassification.forward.
        # Please refer to original code of transformers for more details.
        """
        logger = logging.get_logger(__name__)

        if input_ids is not None:
            batch_size, _ = input_ids.shape[:2]
        else:
            batch_size, _ = hidden_states.shape[:2]
        assert (
            self.config.pad_token_id is not None or batch_size == 1
        ), "Cannot handle batch sizes > 1 if no padding token is defined."

        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        transformer_outputs = GPTJPipelineForwards.gptj_model_forward(
            self.transformer,
            input_ids,
            past_key_values=past_key_values,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            stage_manager=stage_manager,
            hidden_states=hidden_states,
            stage_index=stage_index,
            shard_config=shard_config,
        )

        # If not at the last stage, return hidden_states as in GPTJModel
        if not stage_manager.is_last_stage():
            return {"hidden_states": transformer_outputs["hidden_states"]}

        hidden_states = transformer_outputs[0]
        logits = self.score(hidden_states)

        if self.config.pad_token_id is None:
            sequence_lengths = -1
        else:
            if input_ids is not None:
                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
                sequence_lengths = sequence_lengths % input_ids.shape[-1]
                sequence_lengths = sequence_lengths.to(logits.device)
            else:
                sequence_lengths = -1
                logger.warning_once(
                    f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
                    "unexpected if using padding tokens in conjunction with `inputs_embeds.`"
                )

        pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]

        loss = None
        if labels is not None:
            labels = labels.to(pooled_logits.device)
            if self.config.problem_type is None:
                if self.num_labels == 1:
                    self.config.problem_type = "regression"
                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
                    self.config.problem_type = "single_label_classification"
                else:
                    self.config.problem_type = "multi_label_classification"

            if self.config.problem_type == "regression":
                loss_fct = MSELoss()
                if self.num_labels == 1:
                    loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
                else:
                    loss = loss_fct(pooled_logits, labels)
            elif self.config.problem_type == "single_label_classification":
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
            elif self.config.problem_type == "multi_label_classification":
                loss_fct = BCEWithLogitsLoss()
                loss = loss_fct(pooled_logits, labels)
        if not return_dict:
            output = (pooled_logits,) + transformer_outputs[1:]
            return ((loss,) + output) if loss is not None else output

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

    @staticmethod
    def gptj_for_question_answering_forward(
        self: GPTJForQuestionAnswering,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        token_type_ids: Optional[torch.LongTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        start_positions: Optional[torch.LongTensor] = None,
        end_positions: Optional[torch.LongTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
        stage_manager: Optional[PipelineStageManager] = None,
        hidden_states: Optional[torch.FloatTensor] = None,
        stage_index: Optional[List[int]] = None,
        shard_config: ShardConfig = None,
    ) -> Union[Dict, Tuple, QuestionAnsweringModelOutput]:
        r"""
        start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for position (index) of the start of the labelled span for computing the token classification loss.
            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
            are not taken into account for computing the loss.
        end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for position (index) of the end of the labelled span for computing the token classification loss.
            Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence
            are not taken into account for computing the loss.

        # This function is modified on the basis of transformers.models.gptj.modeling_gptj.GPTJForQuestionAnswering.forward.
        # Please refer to original code of transformers for more details.

        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = GPTJPipelineForwards.gptj_model_forward(
            self.transformer,
            input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            stage_manager=stage_manager,
            hidden_states=hidden_states,
            stage_index=stage_index,
            shard_config=shard_config,
        )

        # If not at the last stage, return hidden_states as in GPTJModel
        if not stage_manager.is_last_stage():
            return {"hidden_states": outputs["hidden_states"]}

        sequence_output = outputs[0]

        logits = self.qa_outputs(sequence_output)
        start_logits, end_logits = logits.split(1, dim=-1)
        start_logits = start_logits.squeeze(-1).contiguous()
        end_logits = end_logits.squeeze(-1).contiguous()

        total_loss = None
        if start_positions is not None and end_positions is not None:
            # If we are on multi-GPU, split add a dimension
            if len(start_positions.size()) > 1:
                start_positions = start_positions.squeeze(-1).to(start_logits.device)
            if len(end_positions.size()) > 1:
                end_positions = end_positions.squeeze(-1).to(end_logits.device)
            # sometimes the start/end positions are outside our model inputs, we ignore these terms
            ignored_index = start_logits.size(1)
            start_positions = start_positions.clamp(0, ignored_index)
            end_positions = end_positions.clamp(0, ignored_index)

            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
            start_loss = loss_fct(start_logits, start_positions)
            end_loss = loss_fct(end_logits, end_positions)
            total_loss = (start_loss + end_loss) / 2

        if not return_dict:
            output = (start_logits, end_logits) + outputs[2:]
            return ((total_loss,) + output) if total_loss is not None else output

        return QuestionAnsweringModelOutput(
            loss=total_loss,
            start_logits=start_logits,
            end_logits=end_logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


def get_gptj_flash_attention_forward():
    from transformers.models.gptj.modeling_gptj import GPTJAttention

    def forward(
        self: GPTJAttention,
        hidden_states: torch.FloatTensor,
        layer_past: Optional[Tuple[torch.Tensor]] = None,
        attention_mask: Optional[dict] = None,
        position_ids: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = False,
        output_attentions: Optional[bool] = False,
    ) -> Union[
        Tuple[torch.Tensor, Tuple[torch.Tensor]],
        Optional[Tuple[torch.Tensor, Tuple[torch.Tensor], Tuple[torch.Tensor, ...]]],
    ]:
        assert head_mask is None, "head_mask is not supported for FlashAttention"
        query = self.q_proj(hidden_states)
        key = self.k_proj(hidden_states)
        value = self.v_proj(hidden_states)

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

        if is_torch_fx_proxy(position_ids) or torch.jit.is_tracing():
            # The logic to conditionally copy to GPU could not be traced, so we do this
            # every time in the torch.fx case
            embed_positions = get_embed_positions(self.embed_positions, position_ids)
        else:
            embed_positions = self._get_embed_positions(position_ids)

        repeated_position_ids = position_ids.unsqueeze(-1).repeat(1, 1, embed_positions.shape[-1])
        sincos = torch.gather(embed_positions, 1, repeated_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)

        if layer_past is not None:
            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)

        if use_cache is True:
            # Note that this cast is quite ugly, but is not implemented before ROPE as the original codebase keeps the key in float32 all along the computation.
            # Reference: https://github.com/kingoflolz/mesh-transformer-jax/blob/f8315e3003033b23f21d78361b288953064e0e76/mesh_transformer/layers.py#L128
            present = (key.to(hidden_states.dtype), value)
        else:
            present = None

        dropout_p = self.attn_dropout.p if self.training else 0.0
        attn_output = ColoAttention.attention(query, key, value, **attention_mask, dropout_p=dropout_p)
        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)
        outputs = (attn_output, present, None)

        return outputs  # a, present, (attentions)

    return forward


def gptj_model_forward_for_flash_attention(shard_config: ShardConfig):
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        token_type_ids: Optional[torch.LongTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        inputs_embeds: Optional[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]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
        elif input_ids is not None:
            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
            input_shape = input_ids.size()
            input_ids = input_ids.view(-1, input_shape[-1])
            input_ids.shape[0]
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
            inputs_embeds.shape[0]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

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

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

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

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

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

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

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

        hidden_states = inputs_embeds

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

        hidden_states = self.drop(hidden_states)

        attention_mask = _get_attention_mask(
            self, shard_config, hidden_states, past_key_values, attention_mask, self._use_flash_attention_2
        )

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

        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

        presents = () if use_cache else None
        all_self_attentions = () if output_attentions else None
        all_hidden_states = () if output_hidden_states else None
        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
            # Model parallel
            if self.model_parallel:
                torch.cuda.set_device(hidden_states.device)
                # Ensure layer_past is on same device as hidden_states (might not be correct)
                if layer_past is not None:
                    layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past)
                # Ensure that attention_mask is always on the same device as hidden_states
                if attention_mask is not None:
                    attention_mask = attention_mask.to(hidden_states.device)
                if isinstance(head_mask, torch.Tensor):
                    head_mask = head_mask.to(hidden_states.device)
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            if self.gradient_checkpointing and self.training:

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

                    return custom_forward

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

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

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

            # Model Parallel: If it's the last layer for that device, put things on the next device
            if self.model_parallel:
                for k, v in self.device_map.items():
                    if i == v[-1] and "cuda:" + str(k) != self.last_device:
                        hidden_states = hidden_states.to("cuda:" + str(k + 1))

        hidden_states = self.ln_f(hidden_states)

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

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

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

    return forward


def gptj_sequence_parallel_forward_fn(shard_config: ShardConfig):
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        token_type_ids: Optional[torch.LongTensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        inputs_embeds: Optional[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]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

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

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

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

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

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

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

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

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

        hidden_states = inputs_embeds

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

        hidden_states = self.drop(hidden_states)

        output_shape = input_shape + (hidden_states.size(-1),)
        attention_mask = _get_attention_mask(
            self, shard_config, hidden_states, past_key_values, attention_mask, self._use_flash_attention_2
        )

        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

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

        # split the input tensor along sequence dimension
        # [batch_size, seq_len, hidden_size] -> [batch_size, seq_len/TP_size, hidden_size]
        hidden_states = split_forward_gather_backward(
            hidden_states,
            dim=1,
            process_group=shard_config.tensor_parallel_process_group,
        )

        for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
            # Model parallel
            if self.model_parallel:
                torch.cuda.set_device(hidden_states.device)
                # Ensure layer_past is on same device as hidden_states (might not be correct)
                if layer_past is not None:
                    layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past)
                # Ensure that attention_mask is always on the same device as hidden_states
                if attention_mask is not None:
                    attention_mask = attention_mask.to(hidden_states.device)
                if isinstance(head_mask, torch.Tensor):
                    head_mask = head_mask.to(hidden_states.device)
            if output_hidden_states:
                all_hidden_states = all_hidden_states + (hidden_states,)

            if self.gradient_checkpointing and self.training:

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

                    return custom_forward

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

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

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

            # Model Parallel: If it's the last layer for that device, put things on the next device
            if self.model_parallel:
                for k, v in self.device_map.items():
                    if i == v[-1] and "cuda:" + str(k) != self.last_device:
                        hidden_states = hidden_states.to("cuda:" + str(k + 1))

        # When sequence parallelism done, gather the output tensor in forward and split it in backward
        hidden_states = gather_forward_split_backward(
            hidden_states,
            dim=1,
            process_group=shard_config.tensor_parallel_process_group,
        )

        hidden_states = self.ln_f(hidden_states)

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

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

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

    return forward