code/flash-linear-attention/fla/models/sse/modeling_sse.py

from __future__ import annotations

import math
import warnings
from dataclasses import dataclass
from typing import TYPE_CHECKING, Optional, Tuple

import torch
import torch.nn as nn
from transformers.modeling_outputs import BaseModelOutputWithPast, MoeCausalLMOutputWithPast
from transformers.modeling_utils import PreTrainedModel
from transformers.utils import logging
from transformers.utils.deprecation import deprecate_kwarg

from fla.layers.attn import Attention
from fla.layers.sse import SSEGLA, SSEGDN
from fla.models.sse.configuration_sse import SSEConfig
from fla.models.utils import Cache, FLAGenerationMixin
from fla.modules import FusedCrossEntropyLoss, FusedLinearCrossEntropyLoss, RMSNorm
from fla.modules import GatedMLP as SSEMLP
from fla.modules.l2warp import l2_warp

if TYPE_CHECKING:
    from transformers.processing_utils import Unpack


try:
    from transformers.modeling_layers import GradientCheckpointingLayer
except ImportError:
    from fla.models.modeling_layers import GradientCheckpointingLayer

logger = logging.get_logger(__name__)


class SSEBlock(GradientCheckpointingLayer):

    def __init__(self, config: SSEConfig, layer_idx: int):
        super().__init__()

        self.config = config
        self.layer_idx = layer_idx

        self.attn_norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
        if config.attn is not None and layer_idx in config.attn['layers']:
            self.attn = Attention(
                hidden_size=config.hidden_size,
                num_heads=config.attn['num_heads'],
                num_kv_heads=config.attn['num_kv_heads'],
                qkv_bias=config.attn['qkv_bias'],
                window_size=config.attn['window_size'],
                rope_theta=config.attn['rope_theta'],
                max_position_embeddings=config.max_position_embeddings,
                layer_idx=layer_idx,
            )
        elif config.linear_attn_type == "gla":
            self.attn = SSEGLA(
                mode=config.attn_mode,
                hidden_size=config.hidden_size,
                expand_v=config.expand_v,
                head_dim=config.head_dim,
                num_heads=config.num_heads,
                num_v_heads=config.num_v_heads,
                use_output_gate=config.use_output_gate,
                use_short_conv=config.use_short_conv,
                conv_size=config.conv_size,
                num_sparse_partition=config.num_sparse_partition,
                num_writer=config.num_writer,
                num_reader=config.num_reader,
                sse_implementation=config.sse_implementation,
                norm_eps=config.norm_eps,
                layer_idx=layer_idx,
            )
        elif config.linear_attn_type == "gdn":
            self.attn = SSEGDN(
                mode=config.attn_mode,
                hidden_size=config.hidden_size,
                expand_v=config.expand_v,
                head_dim=config.head_dim,
                num_heads=config.num_heads,
                num_v_heads=config.num_v_heads,
                use_output_gate=config.use_output_gate,
                use_short_conv=config.use_short_conv,
                allow_neg_eigval=config.allow_neg_eigval,
                conv_size=config.conv_size,
                num_sparse_partition=config.num_sparse_partition,
                num_writer=config.num_writer,
                num_reader=config.num_reader,
                sse_implementation=config.sse_implementation,
                norm_eps=config.norm_eps,
                layer_idx=layer_idx,
            )
        else:
            raise ValueError(f"Unknown linear attention type: {config.linear_attn_type}")
        self.mlp_norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)
        self.mlp = SSEMLP(
            hidden_size=config.hidden_size,
            hidden_ratio=config.hidden_ratio,
            intermediate_size=config.intermediate_size,
            hidden_act=config.hidden_act,
            fuse_swiglu=config.fuse_swiglu,
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        past_key_values: Cache | list[torch.FloatTensor] | None = None,
        use_cache: bool | None = False,
        output_attentions: bool | None = False,
        **kwargs: Unpack[dict],
    ) -> tuple[torch.FloatTensor, tuple[torch.FloatTensor, torch.FloatTensor] | None]:
        residual = hidden_states
        hidden_states = self.attn_norm(hidden_states)
        hidden_states, attentions, past_key_values = self.attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            use_cache=use_cache,
            output_attentions=output_attentions,
            **kwargs,
        )
        if self.config.fuse_norm:
            hidden_states, residual = self.mlp_norm(hidden_states, residual, True)
        else:
            hidden_states = residual + hidden_states
            residual = hidden_states
            hidden_states = self.mlp_norm(hidden_states)
        hidden_states = self.mlp(hidden_states, **kwargs)
        hidden_states = residual + hidden_states

        aux_loss = torch.zeros(()).to(hidden_states)
        # Compatible with Attention output
        if isinstance(attentions, tuple):
            attentions, aux_loss = attentions

        outputs = (hidden_states, attentions, past_key_values, aux_loss)

        return outputs


class SSEPreTrainedModel(PreTrainedModel):

    config_class = SSEConfig
    base_model_prefix = 'model'
    supports_gradient_checkpointing = True
    _no_split_modules = ['SSEBlock']
    _supports_cache_class = True

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

    def _init_weights(
        self,
        module: nn.Module,
        prenorm_residual_strategy: str | None = None,
        num_residuals_per_layer: int = 2,
    ):
        if isinstance(module, SSEGDN) and next(module.parameters()).device.type != 'meta':
            with torch.no_grad():
                module.A_log.copy_(nn.init.uniform_(module.A_log, a=0, b=16).log())
                module.A_log._no_weight_decay = True
                dt = torch.exp(
                    nn.init.uniform_(module.dt_bias) * (math.log(0.1) - math.log(0.001)) + math.log(0.001),
                ).clamp(min=1e-4)
                # Inverse of softplus: https://github.com/pytorch/pytorch/issues/72759
                inv_dt = dt + torch.log(-torch.expm1(-dt))
                module.dt_bias.copy_(inv_dt)
                module.dt_bias._no_weight_decay = True

        elif isinstance(module, (nn.Linear, nn.Conv1d)):
            # Slightly different from the TF version which uses truncated_normal for initialization
            # cf https://github.com/pytorch/pytorch/pull/5617
            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
            if module.bias is not None:
                nn.init.zeros_(module.bias)
        elif isinstance(module, nn.Embedding):
            nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
        elif hasattr(module, 'reset_parameters'):
            module.reset_parameters()

        if prenorm_residual_strategy is not None:
            # Reinitialize selected weights subject to the OpenAI GPT-2 Paper Scheme:
            #   > A modified initialization which accounts for the accumulation on the residual path with model depth. Scale
            #   > the weights of residual layers at initialization by a factor of 1/√N where N is the # of residual layers.
            #   >   -- GPT-2 :: https://openai.com/blog/better-language-models/
            #
            # Reference (Megatron-LM): https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/model/gpt_model.py
            p = None
            if hasattr(module, 'o_proj'):
                p = module.o_proj.weight
            elif hasattr(module, 'down_proj'):
                p = module.down_proj.weight
            if p is not None:
                # Special Scaled Initialization --> There are 2 Layer Norms per Transformer Block
                # Following Pytorch init, except scale by 1/sqrt(2 * n_layer)
                # We need to reinit p since this code could be called multiple times
                # Having just p *= scale would repeatedly scale it down
                if prenorm_residual_strategy == 'rescale':
                    nn.init.kaiming_uniform_(p, a=math.sqrt(5))
                    with torch.no_grad():
                        p /= math.sqrt(num_residuals_per_layer * self.config.num_hidden_layers)
                elif prenorm_residual_strategy == 'zero':
                    nn.init.zeros_(p)
                else:
                    raise ValueError(f"Invalid prenorm_residual_strategy: {prenorm_residual_strategy}")


@dataclass
class MoeModelOutputWithPastAndAuxLosses(BaseModelOutputWithPast):
    """
    Base class for model's outputs, with potential hidden states and attentions.

    Args:
        aux_losses (`Optional[Tuple[torch.FloatTensor]]`, *optional*, returned when `labels` is provided):
            aux_losses for the sparse modules.
    """

    aux_losses: Optional[Tuple[torch.FloatTensor]] = None


class SSEModel(SSEPreTrainedModel):

    def __init__(self, config: SSEConfig):
        super().__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
        self.layers = nn.ModuleList([SSEBlock(config, layer_idx) for layer_idx in range(config.num_hidden_layers)])
        self.norm = (RMSNorm if config.fuse_norm else nn.RMSNorm)(config.hidden_size, eps=config.norm_eps)

        self.gradient_checkpointing = False

        self.post_init()

    def get_input_embeddings(self):
        return self.embeddings

    def set_input_embeddings(self, value):
        self.embeddings = value

    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: Optional[torch.Tensor] = None,  # noqa
        inputs_embeds: torch.FloatTensor | None = None,
        past_key_values: Cache | list[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: Unpack[dict],
    ) -> tuple | BaseModelOutputWithPast:
        if output_attentions:
            warnings.warn("`SSEModel` does not `output_attentions` now, setting it to `False`.")
            output_attentions = False
        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
        output_aux_losses = True
        use_cache = use_cache if use_cache is not None else (self.config.use_cache if not self.training else False)
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # retrieve input_ids and inputs_embeds
        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")
        if input_ids is None and inputs_embeds is None:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        if inputs_embeds is None:
            inputs_embeds = self.embeddings(input_ids)
        hidden_states = inputs_embeds

        if use_cache and not isinstance(past_key_values, Cache):
            past_key_values = Cache.from_legacy_cache(past_key_values)

        all_hidden_states = () if output_hidden_states else None
        all_attns = () if output_attentions else None
        all_aux_losses = () if output_aux_losses else None
        for layer in self.layers:
            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            hidden_states, attentions, past_key_values, aux_loss = layer(
                hidden_states,
                attention_mask=attention_mask,
                past_key_values=past_key_values,
                use_cache=use_cache,
                output_attentions=output_attentions,
                **kwargs,
            )

            if output_attentions:
                all_attns += (attentions,)
            
            if output_aux_losses:
                all_aux_losses += (aux_loss,)

        hidden_states = self.norm(hidden_states)

        # add hidden states from the last decoder layer
        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        if not return_dict:
            return tuple(i for i in [hidden_states, past_key_values, all_hidden_states, all_attns, all_aux_losses] if i is not None)
        return MoeModelOutputWithPastAndAuxLosses(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
            hidden_states=all_hidden_states,
            attentions=all_attns,
            aux_losses=all_aux_losses,
        )


class SSEForCausalLM(SSEPreTrainedModel, FLAGenerationMixin):

    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.model = SSEModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        self.criterion = None
        self.aux_loss_coef = config.aux_loss_coef

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

    def get_input_embeddings(self):
        return self.model.embeddings

    def set_input_embeddings(self, value):
        self.model.embeddings = value

    def get_output_embeddings(self):
        return self.lm_head

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

    def set_decoder(self, decoder):
        self.model = decoder

    def get_decoder(self):
        return self.model

    def generate(self, *args, **kwargs):
        try:
            return super().generate(*args, **kwargs)
        except AttributeError as exception:
            if 'past_key_values' in str(exception):
                raise AttributeError(
                    f"You tried to call `generate` with a decoding strategy that manipulates `past_key_values`, "
                    f"which is not supported for {self.__class__.__name__}. "
                    f"Try another generation strategy instead. "
                    f"For the available generation strategies, check this doc: "
                    f"https://huggingface.co/docs/transformers/en/generation_strategies#decoding-strategies",
                )
            else:
                raise exception

    @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: torch.Tensor | None = None,
        inputs_embeds: torch.Tensor | None = None,
        past_key_values: Cache | list[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 | None = 0,
        **kwargs: Unpack[dict],
    ) -> tuple | MoeCausalLMOutputWithPast:
        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

        outputs = self.model(
            input_ids=input_ids,
            attention_mask=attention_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,
            **kwargs,
        )

        hidden_states = outputs[0]

        loss, aux_loss, logits = None, None, None
        if not self.config.fuse_linear_cross_entropy or labels is None:
            logits = self.lm_head(hidden_states if logits_to_keep is None else hidden_states[:, -logits_to_keep:])
        if labels is not None:
            if getattr(self, 'criterion', None) is None:
                if self.config.fuse_linear_cross_entropy:
                    criterion = FusedLinearCrossEntropyLoss(use_l2warp=self.config.use_l2warp)
                elif self.config.fuse_cross_entropy:
                    criterion = FusedCrossEntropyLoss(inplace_backward=True)
                else:
                    criterion = nn.CrossEntropyLoss()
            else:
                criterion = self.criterion
            labels = labels.to(hidden_states.device)
            labels = torch.cat((labels[..., 1:], torch.full_like(labels[:, :1], criterion.ignore_index)), 1)
            if self.config.fuse_linear_cross_entropy:
                loss = criterion(hidden_states, labels, self.lm_head.weight, self.lm_head.bias)
            else:
                loss = criterion(logits.view(labels.numel(), -1), labels.view(-1))
                loss = l2_warp(loss, logits) if self.config.use_l2warp else loss
            
            aux_losses = outputs.aux_losses
            compute_device = aux_losses[0].device
            aux_loss = sum(layer_aux_loss.to(compute_device) for layer_aux_loss in aux_losses)

            loss += self.aux_loss_coef * aux_loss.to(loss.device)

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

        return MoeCausalLMOutputWithPast(
            loss=loss,
            aux_loss=aux_loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )