modeling_aliceai_t5.py

from collections.abc import Callable
from typing import Optional

import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers.cache_utils import Cache, DynamicCache, EncoderDecoderCache, StaticCache
from transformers.configuration_utils import PretrainedConfig
from transformers.generation.utils import GenerationMixin
from transformers.masking_utils import (
    create_bidirectional_mask,
    create_bidirectional_sliding_window_mask,
    create_causal_mask,
    create_sliding_window_causal_mask,
)
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
from transformers.modeling_layers import GradientCheckpointingLayer
from transformers.modeling_outputs import (
    BaseModelOutput,
    BaseModelOutputWithPastAndCrossAttentions,
    Seq2SeqLMOutput,
    Seq2SeqModelOutput,
)
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from transformers.processing_utils import Unpack
from transformers.utils import logging
from transformers.utils.generic import can_return_tuple

from .configuration_aliceai_t5 import AliceAIT5Config, AliceAIT5ModuleConfig

logger = logging.get_logger(__name__)


class AliceAIT5RMSNorm(nn.Module):
    def __init__(self, dim: int, eps: float = 1e-6):
        super().__init__()
        self.eps = eps
        self.weight = nn.Parameter(torch.zeros(dim))

    def forward(self, x):
        return F.rms_norm(x.to(self.weight.dtype), x.shape[-1:], self.weight, self.eps)

    def extra_repr(self):
        return f"{tuple(self.weight.shape)}, eps={self.eps}"


class AliceAIT5RotaryEmbedding(nn.Module):
    def __init__(self, config, device=None):
        super().__init__()
        rope_type = config.rope_parameters.get("rope_type", "default")
        if "dynamic" in rope_type or rope_type == "longrope":
            raise ValueError(f"{rope_type} requires dynamic RoPE updates; this model uses fixed rotary tables.")

        self.config = config
        self._rope_initialized = False
        self.rope_init_fn = self.compute_default_rope_parameters
        if rope_type != "default":
            self.rope_init_fn = ROPE_INIT_FUNCTIONS[rope_type]

        inv_freq, attention_scaling = self.rope_init_fn(self.config, device)
        self._build_cos_sin(inv_freq, attention_scaling)
        if device is not None and str(device) != "meta":
            self._rope_initialized = True

    def _build_cos_sin(self, inv_freq: torch.Tensor, attention_scaling: float):
        positions = torch.arange(
            self.config.max_position_embeddings,
            dtype=torch.float32,
            device=inv_freq.device,
        )
        angles = torch.outer(positions, inv_freq.float())
        cos = angles.cos() * attention_scaling
        sin = angles.sin() * attention_scaling
        self.register_buffer("cos", cos, persistent=False)
        self.register_buffer("sin", sin, persistent=False)

    @staticmethod
    def compute_default_rope_parameters(
        config: PretrainedConfig | None = None,
        device: Optional["torch.device"] = None,
        seq_len: int | None = None,
    ) -> tuple["torch.Tensor", float]:
        """Return default RoPE inverse frequencies; ``seq_len`` is unused."""
        base = config.rope_parameters["rope_theta"]
        partial_rotary_factor = config.rope_parameters.get("partial_rotary_factor", 1.0)
        head_dim = getattr(config, "head_dim", None) or config.hidden_size // config.num_attention_heads
        dim = int(head_dim * partial_rotary_factor)

        inv_freq = 1.0 / (
            base ** (torch.arange(0, dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / dim)
        )
        return inv_freq, 1.0

    @torch.no_grad()
    def forward(self, x, position_ids):
        if self.cos.device.type == "meta" or not self._rope_initialized:
            inv_freq, attention_scaling = self.rope_init_fn(self.config, x.device)
            self._build_cos_sin(inv_freq, attention_scaling)
            self._rope_initialized = True
        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
        with torch.autocast(device_type=device_type, enabled=False):
            return self.cos[position_ids].to(x.dtype), self.sin[position_ids].to(x.dtype)


def rotate_half_torch(x):
    x1, x2 = x.chunk(2, dim=-1)
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_emb_torch(x, cos, sin):
    """
    x: (batch_size, seqlen, nheads, headdim)
    cos, sin: (seqlen, rotary_dim / 2) or (batch_size, seqlen, rotary_dim / 2)
    """
    ro_dim = cos.shape[-1] * 2
    assert ro_dim <= x.shape[-1]
    cos = torch.cat((cos, cos), dim=-1)
    sin = torch.cat((sin, sin), dim=-1)
    cos = cos.unsqueeze(-2)
    sin = sin.unsqueeze(-2)
    return torch.cat(
        [x[..., :ro_dim] * cos + rotate_half_torch(x[..., :ro_dim]) * sin, x[..., ro_dim:]],
        dim=-1,
    )


def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """Repeat key/value heads to match the number of query heads."""
    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
    if n_rep == 1:
        return hidden_states
    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)


def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: torch.Tensor | None,
    dropout: float = 0.0,
    scaling: float | None = None,
    softcap: float | None = None,
    **kwargs,
) -> tuple[torch.Tensor, torch.Tensor]:
    if scaling is None:
        scaling = module.head_dim**-0.5

    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)

    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling

    if softcap is not None:
        attn_weights = attn_weights / softcap
        attn_weights = torch.tanh(attn_weights)
        attn_weights = attn_weights * softcap
    if attention_mask is not None:
        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
        attn_weights = attn_weights + causal_mask

    attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
    attn_weights = F.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value_states)
    attn_output = attn_output.transpose(1, 2).contiguous()
    return attn_output, attn_weights


class AliceAIT5SelfAttention(nn.Module):
    """Self-attention with rotary position embeddings and grouped key/value heads."""

    def __init__(self, config: AliceAIT5ModuleConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
        self.scaling = config.query_pre_attn_scalar**-0.5
        self.attention_dropout = self.config.attention_dropout
        # FlashAttention reads causality from the module.
        self.is_causal = config.is_decoder

        self.q_proj = nn.Linear(
            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
        )
        self.k_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.v_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.o_proj = nn.Linear(
            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
        )
        self.attn_logit_softcapping = self.config.attn_logit_softcapping
        self.sliding_window = config.sliding_window if config.layer_types[layer_idx] == "sliding_attention" else None

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: torch.Tensor | None,
        past_key_value: Cache | None = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)

        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)

        cos, sin = position_embeddings
        query_states = apply_rotary_emb_torch(query_states.transpose(1, 2), cos, sin).transpose(1, 2)
        key_states = apply_rotary_emb_torch(key_states.transpose(1, 2), cos, sin).transpose(1, 2)

        if past_key_value is not None:
            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx)

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )

        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            dropout=self.attention_dropout if self.training else 0.0,
            scaling=self.scaling,
            sliding_window=self.sliding_window,
            softcap=self.attn_logit_softcapping,
            **kwargs,
        )
        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights


class AliceAIT5CrossAttention(nn.Module):
    """Non-causal decoder attention over cached encoder keys and values."""

    def __init__(self, config: AliceAIT5ModuleConfig, layer_idx: int):
        super().__init__()
        if config.cross_attention_hidden_size is None:
            raise ValueError("Cross-attention needs cross_attention_hidden_size to be specified.")

        self.config = config
        self.layer_idx = layer_idx
        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
        self.scaling = config.query_pre_attn_scalar**-0.5
        self.attention_dropout = self.config.attention_dropout

        self.is_causal = False

        self.q_proj = nn.Linear(
            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
        )

        self.k_proj = nn.Linear(
            config.cross_attention_hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.v_proj = nn.Linear(
            config.cross_attention_hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.o_proj = nn.Linear(
            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
        )
        self.attn_logit_softcapping = self.config.attn_logit_softcapping

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor | None,
        encoder_hidden_states: torch.Tensor | None,
        past_key_value: EncoderDecoderCache | None = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        if encoder_hidden_states is None:
            raise ValueError("Encoder hidden state is required for cross attention.")

        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)
        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)

        if past_key_value is not None:
            is_updated = past_key_value.is_updated.get(self.layer_idx)
            curr_past_key_value = past_key_value.cross_attention_cache

        if past_key_value is None or not is_updated:
            encoder_input_shape = encoder_hidden_states.shape[:-1]
            encoder_hidden_shape = (*encoder_input_shape, -1, self.head_dim)
            key_states = self.k_proj(encoder_hidden_states).view(encoder_hidden_shape).transpose(1, 2)
            value_states = self.v_proj(encoder_hidden_states).view(encoder_hidden_shape).transpose(1, 2)

            if past_key_value is not None:
                key_states, value_states = curr_past_key_value.update(key_states, value_states, self.layer_idx)
                past_key_value.is_updated[self.layer_idx] = True
        else:
            key_states = curr_past_key_value.layers[self.layer_idx].keys
            value_states = curr_past_key_value.layers[self.layer_idx].values

        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
            self.config._attn_implementation, eager_attention_forward
        )

        # Pack valid encoder keys/values separately from decoder queries so
        # right padding cannot mask out decoder positions.
        extra_attention_kwargs = {}
        if (
            self.config._attn_implementation.startswith("flash_attention")
            and attention_mask is not None
            and attention_mask.ndim == 2
        ):
            batch_size, query_length = hidden_states.shape[:2]
            key_mask = attention_mask.to(device=key_states.device, dtype=torch.bool)
            key_states = key_states.transpose(1, 2)[key_mask].transpose(0, 1).unsqueeze(0)
            value_states = value_states.transpose(1, 2)[key_mask].transpose(0, 1).unsqueeze(0)
            query_states = (
                query_states.transpose(1, 2)
                .reshape(1, batch_size * query_length, self.config.num_attention_heads, self.head_dim)
                .transpose(1, 2)
            )

            key_lengths = key_mask.sum(dim=-1, dtype=torch.int32)
            cu_seq_lens_k = F.pad(key_lengths.cumsum(dim=0, dtype=torch.int32), (1, 0))
            cu_seq_lens_q = torch.arange(
                0,
                (batch_size + 1) * query_length,
                query_length,
                dtype=torch.int32,
                device=hidden_states.device,
            )
            extra_attention_kwargs = {
                "cu_seq_lens_q": cu_seq_lens_q,
                "cu_seq_lens_k": cu_seq_lens_k,
                "max_length_q": query_length,
                "max_length_k": int(key_lengths.max().item()),
            }
            attention_mask = None

        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            dropout=self.attention_dropout if self.training else 0.0,
            scaling=self.scaling,
            sliding_window=None,
            softcap=self.attn_logit_softcapping,
            **extra_attention_kwargs,
            **kwargs,
        )

        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights


class AliceAIT5EncoderLayer(GradientCheckpointingLayer):
    """Encoder sub-layer."""

    def __init__(self, config, layer_idx: int, *, mlp: nn.Module):
        super().__init__()
        self.config = config
        self.attention_type = config.layer_types[layer_idx]

        self.pre_self_attn_layernorm = AliceAIT5RMSNorm(config.hidden_size, eps=config.norm_eps)
        self.self_attn = AliceAIT5SelfAttention(config=config, layer_idx=layer_idx)
        self.post_self_attn_layernorm = AliceAIT5RMSNorm(config.hidden_size, eps=config.norm_eps)
        self.mlp = mlp

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        output_attentions: bool | None = False,
        **kwargs,
    ) -> tuple[
        torch.FloatTensor,
        tuple[torch.FloatTensor, torch.FloatTensor] | None,
    ]:
        mlp_attn_dtype = self.self_attn.q_proj.weight.dtype

        residual = hidden_states
        hidden_states = self.pre_self_attn_layernorm(hidden_states)
        hidden_states, self_attn_weights = self.self_attn(
            hidden_states=hidden_states.to(mlp_attn_dtype),
            position_embeddings=position_embeddings,
            attention_mask=attention_mask,
            position_ids=position_ids,
            output_attentions=output_attentions,
            use_cache=False,
            past_key_value=None,
            **kwargs,
        )
        if self.config.fp32_residual:
            hidden_states = residual.float() + hidden_states.float()
        else:
            hidden_states = residual + hidden_states

        residual = hidden_states

        hidden_states = self.post_self_attn_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states.to(mlp_attn_dtype))
        if self.config.fp32_residual:
            hidden_states = residual.float() + hidden_states.float()
        else:
            hidden_states = residual + hidden_states

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights,)

        return outputs


class AliceAIT5DecoderLayer(AliceAIT5EncoderLayer):
    """Decoder sub-layer: an extra cross-attention layer."""

    def __init__(self, config, layer_idx: int, *, mlp: nn.Module):
        super().__init__(config, layer_idx, mlp=mlp)
        self.cross_attn = AliceAIT5CrossAttention(config=config, layer_idx=layer_idx)
        self.post_cross_attn_layernorm = AliceAIT5RMSNorm(config.hidden_size, eps=config.norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_value: EncoderDecoderCache | None = None,
        output_attentions: bool | None = False,
        use_cache: bool | None = False,
        encoder_hidden_states: torch.Tensor | None = None,
        encoder_attention_mask: torch.Tensor | None = None,
        **kwargs,
    ) -> tuple[
        torch.FloatTensor,
        tuple[torch.FloatTensor, torch.FloatTensor] | None,
        tuple[torch.FloatTensor, torch.FloatTensor] | None,
    ]:
        mlp_attn_dtype = self.self_attn.q_proj.weight.dtype
        residual = hidden_states
        hidden_states = self.pre_self_attn_layernorm(hidden_states)
        hidden_states, self_attn_weights = self.self_attn(
            hidden_states=hidden_states.to(mlp_attn_dtype),
            position_embeddings=position_embeddings,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value.self_attention_cache if past_key_value is not None else None,
            output_attentions=output_attentions,
            use_cache=use_cache,
            **kwargs,
        )

        if self.config.fp32_residual:
            hidden_states = residual.float() + hidden_states.float()
        else:
            hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.post_self_attn_layernorm(hidden_states)

        hidden_states, cross_attn_weights = self.cross_attn(
            hidden_states=hidden_states.to(mlp_attn_dtype),
            encoder_hidden_states=encoder_hidden_states.to(mlp_attn_dtype),
            attention_mask=encoder_attention_mask,
            past_key_value=past_key_value,
            output_attentions=output_attentions,
            use_cache=use_cache,
            **kwargs,
        )

        if self.config.fp32_residual:
            hidden_states = residual.float() + hidden_states.float()
        else:
            hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.post_cross_attn_layernorm(hidden_states)

        hidden_states = self.mlp(hidden_states.to(mlp_attn_dtype))
        if self.config.fp32_residual:
            hidden_states = residual.float() + hidden_states.float()
        else:
            hidden_states = residual + hidden_states

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights, cross_attn_weights)

        return outputs


class AliceAIT5LMHead(nn.Module):
    """Head for language modeling (generation) tasks."""

    def __init__(self, hidden_size: int, vocab_size: int, bias: bool = False, dtype: torch.dtype = torch.float):
        super().__init__()
        self.out_proj = nn.Linear(hidden_size, vocab_size, bias=bias, dtype=dtype)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        # Accumulate low-precision CUDA products into FP32 without copying tied weights.
        output_dtype = torch.float32 if hidden_states.dtype in (torch.float16, torch.bfloat16) else hidden_states.dtype
        flat_states = hidden_states.reshape(-1, hidden_states.shape[-1])
        weight = self.out_proj.weight
        if (
            not torch.is_grad_enabled()
            and flat_states.device.type == "cuda"
            and flat_states.dtype in (torch.float16, torch.bfloat16)
            and weight.dtype == flat_states.dtype
        ):
            logits = torch.mm(flat_states, weight.t(), out_dtype=output_dtype)
        else:
            logits = F.linear(
                flat_states.to(output_dtype),
                weight.to(output_dtype),
            )
        if self.out_proj.bias is not None:
            logits = logits + self.out_proj.bias.to(output_dtype)
        return logits.view(*hidden_states.shape[:-1], weight.shape[0])


class AliceAIT5PreTrainedModel(PreTrainedModel):
    config_class = AliceAIT5Config
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = []
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn = True
    _supports_attention_backend = True

    def gradient_checkpointing_enable(self, gradient_checkpointing_kwargs=None):
        from .moe_layers import _checkpoint_contexts

        options = dict(gradient_checkpointing_kwargs or {})
        if options.get("use_reentrant", False):
            raise ValueError("AliceAIT5 gradient checkpointing requires use_reentrant=False.")
        if "context_fn" in options:
            raise ValueError(
                "Custom checkpointing context_fn is not supported; router statistics need recompute control."
            )
        if options.get("debug", False):
            raise ValueError("Checkpointing debug=True is incompatible with router recomputation contexts.")
        options["use_reentrant"] = False
        options["context_fn"] = _checkpoint_contexts
        return super().gradient_checkpointing_enable(gradient_checkpointing_kwargs=options)

    @classmethod
    def from_pretrained(cls, *args, **kwargs):
        if any(kwargs.get(name) is not None for name in ("tp_plan", "tp_size", "device_mesh", "distributed_config")):
            raise ValueError("Tensor/expert parallelism is not implemented for AliceAIT5; omit TP arguments.")
        return super().from_pretrained(*args, **kwargs)

    def resize_token_embeddings(
        self,
        new_num_tokens: int | None = None,
        pad_to_multiple_of: int | None = None,
        mean_resizing: bool = True,
    ) -> nn.Embedding:
        model_embeds = super().resize_token_embeddings(
            new_num_tokens=new_num_tokens,
            pad_to_multiple_of=pad_to_multiple_of,
            mean_resizing=mean_resizing,
        )
        vocab_size = model_embeds.weight.shape[0]
        self.config.vocab_size = vocab_size
        for subconfig_name in ("encoder", "decoder"):
            subconfig = getattr(self.config, subconfig_name, None)
            if subconfig is not None:
                subconfig.vocab_size = vocab_size
        if not getattr(self.config, "shared_embeddings", True) and (
            new_num_tokens is not None or pad_to_multiple_of is not None
        ):
            decoder = self.get_decoder()
            if decoder is not self:
                decoder.resize_token_embeddings(
                    vocab_size,
                    mean_resizing=mean_resizing,
                )
                self.tie_weights()
        return model_embeds

    def _init_weights(self, module):
        std = self.config.initializer_range
        if isinstance(module, nn.Linear):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, AliceAIT5RMSNorm):
            module.weight.data.fill_(1.0)
        elif isinstance(module, AliceAIT5LMHead):
            if not self.config.tie_word_embeddings:
                scale = module.out_proj.weight.shape[0] ** -0.5
                module.out_proj.weight.data.normal_(mean=0.0, std=std * scale)

    def _shift_right(self, input_ids):
        """Prepend decoder BOS and replace ignored labels with the padding token."""
        decoder_start_token_id = self.config.decoder.bos_token_id
        pad_token_id = self.config.decoder.pad_token_id

        if decoder_start_token_id is None:
            raise ValueError("self.model.config.decoder.bos_token_id has to be defined. ")

        shifted_input_ids = input_ids.new_zeros(input_ids.shape)
        shifted_input_ids[..., 1:] = input_ids[..., :-1].clone()
        shifted_input_ids[..., 0] = decoder_start_token_id

        if pad_token_id is None:
            raise ValueError("self.model.config.decoder.pad_token_id has to be defined.")

        shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)

        return shifted_input_ids


def make_default_2d_attention_mask(
    token_ids: torch.LongTensor | None,
    hidden_states: torch.Tensor,
    pad_token_id: int | None,
) -> torch.Tensor:
    """Construct the default attention mask."""
    if token_ids is not None:
        if pad_token_id is None:
            raise ValueError("`pad_token_id` is required for padding information.")
        attention_mask = (token_ids != pad_token_id).to(hidden_states.device, torch.long)
    else:
        attention_mask = torch.ones(
            (hidden_states.shape[0], hidden_states.shape[1]), device=hidden_states.device, dtype=torch.long
        )
    return attention_mask


def validate_encoder_attention_mask(attention_mask: torch.Tensor | dict | None):
    if isinstance(attention_mask, torch.Tensor) and attention_mask.ndim == 2:
        if attention_mask.shape[-1] == 0 or not attention_mask.bool().any(dim=-1).all():
            raise ValueError("Each encoder sequence must contain at least one unmasked token.")


class AliceAIT5Encoder(AliceAIT5PreTrainedModel):
    _no_split_modules = [AliceAIT5EncoderLayer.__name__]

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

    def _init_components(self, config):
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size
        self.has_embeddings = config.has_embeddings

        if not config.is_decoder and not config.has_embeddings:
            raise ValueError("Encoder must have embeddings, but got has_embeddings=False with is_decoder=False")

        if config.has_embeddings:
            self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)

        self.norm = AliceAIT5RMSNorm(config.hidden_size, eps=config.norm_eps)

        self.rotary_emb = AliceAIT5RotaryEmbedding(config=config)
        self.gradient_checkpointing = False
        if config.is_decoder:
            self.dropout = nn.Dropout(config.dropout_rate)
        self._build_layers(config)

    def _build_layers(self, config):
        raise NotImplementedError("Use AliceAIT5MoEEncoder to construct encoder layers.")

    def get_input_embeddings(self):
        if not self.has_embeddings:
            raise NotImplementedError("Module has no `embed_tokens` due to config")
        return self.embed_tokens

    def set_input_embeddings(self, value):
        if not self.has_embeddings:
            raise NotImplementedError("Module can't have `embed_tokens` due to config")
        self.embed_tokens = value

    @can_return_tuple
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
    ) -> BaseModelOutput:
        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
        )

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

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

        if position_ids is None:
            position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device).unsqueeze(0)

        if attention_mask is None:
            attention_mask = make_default_2d_attention_mask(input_ids, inputs_embeds, self.config.pad_token_id)

        validate_encoder_attention_mask(attention_mask)

        if not isinstance(self_attn_mask_mapping := attention_mask, dict):
            mask_kwargs = {
                "config": self.config,
                "inputs_embeds": inputs_embeds,
                "attention_mask": attention_mask,
            }
            self_attn_mask_mapping = {
                "full_attention": create_bidirectional_mask(**mask_kwargs),
            }
            if self.config.sliding_window is not None:
                self_attn_mask_mapping["sliding_attention"] = create_bidirectional_sliding_window_mask(**mask_kwargs)

        hidden_states = inputs_embeds

        position_embeddings = self.rotary_emb(hidden_states, position_ids)

        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None

        for layer_module in self.layers[: self.config.num_hidden_layers]:
            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            layer_outputs = layer_module(
                hidden_states,
                position_embeddings,
                self_attn_mask_mapping[layer_module.attention_type],
                position_ids,
                output_attentions,
                **flash_attn_kwargs,
            )

            hidden_states = layer_outputs[0]

            if output_attentions:
                all_self_attns += (layer_outputs[1],)

        if self.norm is not None:
            hidden_states = self.norm(hidden_states)
        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        return BaseModelOutput(
            last_hidden_state=hidden_states,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
        )


class AliceAIT5Decoder(AliceAIT5Encoder):
    _no_split_modules = [AliceAIT5DecoderLayer.__name__]

    def _build_layers(self, config):
        raise NotImplementedError("Use AliceAIT5MoEDecoder to construct decoder layers.")

    @can_return_tuple
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.Tensor | None = None,
        position_ids: torch.LongTensor | None = None,
        past_key_values: EncoderDecoderCache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        use_cache: bool | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        encoder_hidden_states: torch.Tensor | None = None,
        encoder_attention_mask: torch.Tensor | None = None,
        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
    ) -> BaseModelOutputWithPastAndCrossAttentions:
        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

        if input_ids is not None and not self.has_embeddings:
            raise ValueError(
                "Cannot process input_ids with has_embeddings=False (decoder has no embeddings when shared_embeddings=True)"
            )

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

        if encoder_hidden_states is None:
            raise ValueError("`encoder_hidden_states` must be given in decoder")

        if (
            past_key_values is not None
            and isinstance(past_key_values.self_attention_cache, StaticCache)
            and self.config._attn_implementation.startswith("flash_attention")
        ):
            raise ValueError(
                "FlashAttention with StaticCache is not supported because unwritten cache capacity cannot be masked safely."
            )

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

        # Build an implicit padding mask before creating a cache. In the
        # shared-embedding path the outer model supplies this mask explicitly.
        if attention_mask is None and past_key_values is None:
            attention_mask = make_default_2d_attention_mask(input_ids, inputs_embeds, self.config.pad_token_id)

        if not self.training and use_cache and past_key_values is None:
            past_key_values = EncoderDecoderCache(
                DynamicCache(config=self.config),
                DynamicCache(),
            )

        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
            ).unsqueeze(0)

        if not isinstance(self_attn_mask_mapping := attention_mask, dict):
            mask_kwargs = {
                "config": self.config,
                "inputs_embeds": inputs_embeds,
                "attention_mask": attention_mask,
                "past_key_values": past_key_values.self_attention_cache if past_key_values is not None else None,
                "position_ids": position_ids,
            }
            self_attn_mask_mapping = {
                "full_attention": create_causal_mask(**mask_kwargs),
            }
            if self.config.sliding_window is not None:
                self_attn_mask_mapping["sliding_attention"] = create_sliding_window_causal_mask(**mask_kwargs)

        if not isinstance(cross_attn_mask_mapping := encoder_attention_mask, dict):
            cross_attn_mask_mapping = {
                "full_attention": create_bidirectional_mask(
                    config=self.config,
                    inputs_embeds=inputs_embeds,
                    attention_mask=encoder_attention_mask,
                    encoder_hidden_states=encoder_hidden_states,
                ),
            }

        hidden_states = inputs_embeds

        position_embeddings = self.rotary_emb(hidden_states, position_ids)

        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None
        all_cross_attns = () if output_attentions else None

        hidden_states = self.dropout(hidden_states)

        for layer_module in self.layers[: self.config.num_hidden_layers]:
            if output_hidden_states:
                all_hidden_states += (hidden_states,)

            layer_outputs = layer_module(
                hidden_states=hidden_states,
                position_embeddings=position_embeddings,
                attention_mask=self_attn_mask_mapping[layer_module.attention_type],
                position_ids=position_ids,
                past_key_value=past_key_values,
                output_attentions=output_attentions,
                use_cache=use_cache,
                encoder_hidden_states=encoder_hidden_states,
                encoder_attention_mask=cross_attn_mask_mapping["full_attention"],
                **flash_attn_kwargs,
            )

            hidden_states = layer_outputs[0]

            if output_attentions:
                all_self_attns += (layer_outputs[1],)
                all_cross_attns += (layer_outputs[2],)

        hidden_states = self.norm(hidden_states)

        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
            cross_attentions=all_cross_attns,
        )


class AliceAIT5Model(AliceAIT5PreTrainedModel):
    _no_split_modules = [AliceAIT5EncoderLayer.__name__, AliceAIT5DecoderLayer.__name__]

    def __init__(self, config: AliceAIT5Config):
        super().__init__(config)

        if not config.is_encoder_decoder:
            raise ValueError(
                "AliceAIT5Model only support encoder-decoder modeling. Use `AliceAIT5EncoderModel` instead."
            )

        self.encoder = AliceAIT5Encoder(config.encoder)
        self.decoder = AliceAIT5Decoder(config.decoder)

        self.post_init()

    def get_encoder(self):
        return self.encoder

    def get_decoder(self):
        return self.decoder

    def get_input_embeddings(self):
        return self.encoder.get_input_embeddings()

    def set_input_embeddings(self, new_embeddings):
        return self.encoder.set_input_embeddings(new_embeddings)

    @can_return_tuple
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        decoder_input_ids: torch.LongTensor | None = None,
        decoder_attention_mask: torch.BoolTensor | None = None,
        decoder_position_ids: torch.LongTensor | None = None,
        encoder_outputs: BaseModelOutput | None = None,
        past_key_values: EncoderDecoderCache | None = None,
        inputs_embeds: torch.Tensor | None = None,
        decoder_inputs_embeds: torch.Tensor | None = None,
        use_cache: bool | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
    ) -> Seq2SeqModelOutput:
        """Encode the input and decode the target, optionally reusing cached states."""
        use_cache = use_cache if use_cache is not None else self.config.use_cache

        # Canonicalize implicit masks while token IDs are still available.
        # The same encoder mask is used for encoder self-attention and decoder
        # cross-attention.
        if attention_mask is None:
            if input_ids is not None:
                if self.config.encoder.pad_token_id is None:
                    raise ValueError("`pad_token_id` is required for padding information.")
                attention_mask = input_ids.ne(self.config.encoder.pad_token_id).long()
            elif encoder_outputs is None and inputs_embeds is not None:
                attention_mask = torch.ones(
                    inputs_embeds.shape[:2],
                    dtype=torch.long,
                    device=inputs_embeds.device,
                )

        if decoder_attention_mask is None and past_key_values is None:
            if decoder_input_ids is not None:
                if self.config.decoder.pad_token_id is None:
                    raise ValueError("`pad_token_id` is required for padding information.")
                decoder_attention_mask = decoder_input_ids.ne(self.config.decoder.pad_token_id).long()
            elif decoder_inputs_embeds is not None:
                decoder_attention_mask = torch.ones(
                    decoder_inputs_embeds.shape[:2],
                    dtype=torch.long,
                    device=decoder_inputs_embeds.device,
                )

        if encoder_outputs is None:
            encoder_outputs = self.encoder(
                input_ids=input_ids,
                attention_mask=attention_mask,
                position_ids=position_ids,
                inputs_embeds=inputs_embeds,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict=True,
                **flash_attn_kwargs,
            )
        else:
            validate_encoder_attention_mask(attention_mask)

        encoder_hidden_states = encoder_outputs.last_hidden_state
        if encoder_hidden_states.shape[1] == 0:
            raise ValueError("Each encoder sequence must contain at least one unmasked token.")

        if (decoder_inputs_embeds is None) and self.config.shared_embeddings:
            decoder_inputs_embeds = self.get_input_embeddings()(decoder_input_ids)

        decoder_outputs = self.decoder(
            input_ids=decoder_input_ids if not self.config.shared_embeddings else None,
            attention_mask=decoder_attention_mask,
            position_ids=decoder_position_ids,
            inputs_embeds=decoder_inputs_embeds,
            past_key_values=past_key_values,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=attention_mask,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=True,
            **flash_attn_kwargs,
        )

        return Seq2SeqModelOutput(
            last_hidden_state=decoder_outputs.last_hidden_state,
            past_key_values=decoder_outputs.past_key_values,
            decoder_hidden_states=decoder_outputs.hidden_states,
            decoder_attentions=decoder_outputs.attentions,
            cross_attentions=decoder_outputs.cross_attentions,
            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
            encoder_hidden_states=encoder_outputs.hidden_states,
            encoder_attentions=encoder_outputs.attentions,
        )


class AliceAIT5EncoderModel(AliceAIT5PreTrainedModel):
    def __init__(self, config: AliceAIT5Config):
        super().__init__(config)

        if config.is_encoder_decoder:
            raise ValueError("AliceAIT5EncoderModel only supports encoder-only model. Use `AliceAIT5Model` instead.")

        self.encoder = self._build_encoder(config)
        self.post_init()

    def _build_encoder(self, config):
        return AliceAIT5Encoder(config.encoder)

    def get_input_embeddings(self):
        return self.encoder.get_input_embeddings()

    def set_input_embeddings(self, new_embeddings):
        return self.encoder.set_input_embeddings(new_embeddings)

    @can_return_tuple
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        inputs_embeds: torch.Tensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
    ) -> BaseModelOutput:
        encoder_outputs = self.encoder(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=True,
            **flash_attn_kwargs,
        )
        return encoder_outputs


class AliceAIT5ForConditionalGeneration(AliceAIT5PreTrainedModel, GenerationMixin):
    _no_split_modules = [AliceAIT5EncoderLayer.__name__, AliceAIT5DecoderLayer.__name__]

    def __init__(self, config: AliceAIT5Config):
        config.is_encoder_decoder = True
        self._tied_weights_keys = {
            "lm_head.out_proj.weight": (
                "model.encoder.embed_tokens.weight"
                if config.shared_embeddings
                else "model.decoder.embed_tokens.weight"
            )
        }
        super().__init__(config)

        self.model = self._build_model(config)
        self.vocab_size = config.encoder.vocab_size
        self.lm_head = AliceAIT5LMHead(config.decoder.hidden_size, self.vocab_size, dtype=self.dtype)
        self.loss_type = "ForMaskedLM"
        self.post_init()

    def _build_model(self, config):
        return AliceAIT5Model(config)

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

    def get_output_embeddings(self):
        return self.lm_head.out_proj

    def get_encoder(self):
        return self.model.encoder

    def get_decoder(self):
        return self.model.decoder

    @can_return_tuple
    def forward(
        self,
        input_ids: torch.LongTensor | None = None,
        attention_mask: torch.FloatTensor | None = None,
        position_ids: torch.LongTensor | None = None,
        decoder_input_ids: torch.LongTensor | None = None,
        decoder_attention_mask: torch.BoolTensor | None = None,
        decoder_position_ids: torch.LongTensor | None = None,
        encoder_outputs: BaseModelOutput | None = None,
        past_key_values: EncoderDecoderCache | None = None,
        inputs_embeds: torch.FloatTensor | None = None,
        decoder_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,
        logits_to_keep: int | torch.Tensor = 0,
        **loss_kwargs,
    ) -> tuple[torch.FloatTensor] | Seq2SeqLMOutput:
        """Return decoder logits and optional cross-entropy loss.

        Labels have shape ``[batch_size, target_length]``; ``-100`` is ignored.
        When labels are supplied, all target logits are computed and decoder
        inputs default to the labels shifted right by one position.
        """

        if labels is not None and decoder_input_ids is None and decoder_inputs_embeds is None:
            decoder_input_ids = self._shift_right(labels)

        decoder_outputs: Seq2SeqModelOutput = self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            decoder_input_ids=decoder_input_ids,
            decoder_attention_mask=decoder_attention_mask,
            decoder_position_ids=decoder_position_ids,
            encoder_outputs=encoder_outputs,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            decoder_inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=True,
            **loss_kwargs,
        )

        hidden_states = decoder_outputs.last_hidden_state
        # Loss needs every target position; generation may request only a suffix.
        if labels is not None:
            slice_indices = slice(None)
        elif isinstance(logits_to_keep, int):
            slice_indices = slice(-logits_to_keep, None)
        else:
            slice_indices = logits_to_keep
        logits = self.lm_head(hidden_states[:, slice_indices, :])
        decoder_config = self.get_decoder().config
        if decoder_config.final_logit_softcapping is not None:
            logits = logits / decoder_config.final_logit_softcapping
            logits = torch.tanh(logits)
            logits = logits * decoder_config.final_logit_softcapping

        loss = None
        if labels is not None:
            # Decoder inputs are already shifted; labels stay aligned with logits.
            loss = self.loss_function(logits, labels, self.vocab_size, **loss_kwargs)

        return Seq2SeqLMOutput(
            loss=loss,
            logits=logits,
            past_key_values=decoder_outputs.past_key_values,
            decoder_hidden_states=decoder_outputs.decoder_hidden_states,
            decoder_attentions=decoder_outputs.decoder_attentions,
            cross_attentions=decoder_outputs.cross_attentions,
            encoder_last_hidden_state=decoder_outputs.encoder_last_hidden_state,
            encoder_hidden_states=decoder_outputs.encoder_hidden_states,
            encoder_attentions=decoder_outputs.encoder_attentions,
        )

    def prepare_decoder_input_ids_from_labels(self, labels: torch.Tensor):
        return self._shift_right(labels)


__all__ = [
    "AliceAIT5Config",
    "AliceAIT5ModuleConfig",
    "AliceAIT5ForConditionalGeneration",
    "AliceAIT5Model",
    "AliceAIT5EncoderModel",
    "AliceAIT5PreTrainedModel",
]