modeling_trida.py

from typing import Callable, Optional, Union
from dataclasses import dataclass

import torch
from torch import nn
import torch.nn.functional as F
from functools import partial

from transformers.activations import ACT2FN
from transformers.cache_utils import Cache, DynamicCache
from transformers.generation import GenerationMixin
from transformers.integrations import use_kernel_forward_from_hub
from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
from transformers.modeling_layers import GradientCheckpointingLayer
from transformers.modeling_outputs import (
    BaseModelOutputWithPast,
    CausalLMOutputWithPast,
)
from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
from transformers.processing_utils import Unpack
from transformers.utils import auto_docstring, can_return_tuple, logging
from .configuration_trida import TridaConfig
from torch.nn.attention.flex_attention import flex_attention, create_block_mask
from einops import rearrange, repeat

logger = logging.get_logger(__name__)


@dataclass
class CausalLMOutputWithPastAndBlockCache(CausalLMOutputWithPast):
    block_past_key_values: Optional[Cache] = None
    computed_attention_mask: Optional[torch.Tensor] = None

@dataclass
class BaseModelOutputWithPastAndBlockCache(BaseModelOutputWithPast):
    block_past_key_values: Optional[Cache] = None
    computed_attention_mask: Optional[torch.Tensor] = None


@torch.compile(fullgraph=True, mode="max-autotune-no-cudagraphs")
def fused_flex_attention(q, k, v, mask=None):
    return flex_attention(q, k, v, block_mask=mask, enable_gqa=True)

def block_diff_mask(b, h, q_idx, kv_idx, block_size=None, n=None):
    """
    Constructs the specialized block diffusion attention mask for training
    composed of three masks:
    - **Block Diagonal Mask (M_BD)**: Self-attention within noised blocks
    - **Offset Block Causal Mask (M_OBC)**: Cross-attention for conditional context
    - **Block Causal Mask (M_BC)**: Attention to update x0

    Args:
        b, h: Batch and head indices (ignored for mask logic).
        q_idx, kv_idx: Query and Key indices.
        seq_len: Total sequence length.
        block_size: Defines the block structure.

    Returns:
        A boolean attention mask.
    """
    # Indicate whether token belongs to xt or x0
    x0_flag_q = (q_idx >= n)
    x0_flag_kv = (kv_idx >= n)

    # Compute block indices
    block_q = torch.where(x0_flag_q == 1,
                        (q_idx - n) // block_size,
                        q_idx // block_size)
    block_kv = torch.where(x0_flag_kv == 1,
                        (kv_idx - n) // block_size,
                        kv_idx // block_size)

    # **1. Block Diagonal Mask (M_BD) **
    block_diagonal = (block_q == block_kv) & (x0_flag_q == x0_flag_kv)

    # **2. Offset Block-Causal Mask (M_OBC) **
    offset_block_causal = (
    (block_q > block_kv)
    & (x0_flag_kv == 1)
    & (x0_flag_q == 0)
    )

    # **3. Block-Causal Mask (M_BC) **
    block_causal = (block_q >= block_kv) & (x0_flag_kv == 1) & (x0_flag_q == 1)

    # **4. Combine Masks **
    return block_diagonal | offset_block_causal | block_causal

def eval_block_diff_mask(q_idx, kv_idx, block_size=None):
    # Compute block indices
    block_q = q_idx // block_size
    block_kv = kv_idx // block_size

    return block_q >= block_kv

class TridaMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, x):
        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
        return down_proj


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


def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
    """Applies Rotary Position Embedding to the query and key tensors.

    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        position_ids (`torch.Tensor`, *optional*):
            Deprecated and unused.
        unsqueeze_dim (`int`, *optional*, defaults to 1):
            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
    Returns:
        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
    """
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed


def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """
    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
    """
    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)


class TridaAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: TridaConfig, 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 = self.head_dim**-0.5
        self.attention_dropout = config.attention_dropout
        self.is_causal = True
        self.q_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=False)
        self.k_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=False)
        self.v_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=False)
        self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=False)
        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: Optional[torch.Tensor],
        past_key_value: Optional[Cache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        update_past_key_values: Optional[bool] = False,
        block_past_key_values: Optional[Cache] = None,
        replace_position: Optional[int] = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
        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, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
        if self.training:
            #split q into two parts
            q_1 = query_states[:,:,:query_states.shape[2]//2]
            q_2 = query_states[:,:,query_states.shape[2]//2:]
            #split k into two parts
            k_1 = key_states[:,:,:key_states.shape[2]//2]
            k_2 = key_states[:,:,key_states.shape[2]//2:]
            q_1, k_1 = apply_rotary_pos_emb(q_1, k_1, cos, sin)
            q_2, k_2 = apply_rotary_pos_emb(q_2, k_2, cos, sin)
            query_states = torch.cat((q_1, q_2), dim=-2)
            key_states = torch.cat((k_1, k_2), dim=-2)
        else:
            query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if block_past_key_values is not None:
            if len(block_past_key_values) <= self.layer_idx:
                cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
                key_states, value_states = block_past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
            else:
                block_cache_key_states = block_past_key_values[self.layer_idx][0]
                block_cache_value_states = block_past_key_values[self.layer_idx][1]
                
                block_cache_key_states[:, :, replace_position:replace_position+key_states.shape[2]] = key_states
                block_cache_value_states[:, :, replace_position:replace_position+value_states.shape[2]] = value_states
                key_states = block_cache_key_states
                value_states = block_cache_value_states

        if past_key_value is not None:
            # sin and cos are specific to RoPE models; cache_position needed for the static cache
            if update_past_key_values:
                cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
                key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
            elif len(past_key_value) > self.layer_idx:
                key_states = torch.cat((past_key_value[self.layer_idx][0], key_states), dim=-2)
                value_states = torch.cat((past_key_value[self.layer_idx][1], value_states), dim=-2)

        if self.training:
            attn_output = fused_flex_attention(query_states, key_states, value_states, mask=attention_mask)
            attn_output = attn_output.transpose(1, 2).contiguous()
        else:
            attention_interface = ALL_ATTENTION_FUNCTIONS["sdpa"]

            attn_output, attn_weights = attention_interface(
                self,
                query_states,
                key_states,
                value_states,
                attention_mask,
                is_causal=False,
                dropout=0.0 if not self.training else self.attention_dropout,
                scaling=self.scaling,
                sliding_window=self.sliding_window,  # main diff with Llama
                **kwargs,
            )

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

@use_kernel_forward_from_hub("RMSNorm")
class TridaRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        RMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        input_dtype = hidden_states.dtype
        hidden_states = hidden_states.to(torch.float32)
        variance = hidden_states.pow(2).mean(-1, keepdim=True)
        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
        return self.weight * hidden_states.to(input_dtype)

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


class TridaDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: TridaConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size

        self.self_attn = TridaAttention(config=config, layer_idx=layer_idx)

        self.mlp = TridaMLP(config)
        self.input_layernorm = TridaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = TridaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.attention_type = config.layer_types[layer_idx]

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Cache] = None,
        use_cache: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
        update_past_key_values: Optional[bool] = False,
        use_block_cache: Optional[bool] = False,
        block_past_key_values: Optional[Cache] = None,
        replace_position: Optional[int] = None,
        **kwargs
    ) -> tuple[torch.Tensor]:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        # Self Attention
        hidden_states = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value,
            use_cache=use_cache,
            cache_position=cache_position,
            position_embeddings=position_embeddings,
            update_past_key_values=update_past_key_values,
            use_block_cache=use_block_cache,
            block_past_key_values=block_past_key_values,
            replace_position=replace_position,
            **kwargs,
        )
        hidden_states = residual + hidden_states

        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states
        return hidden_states



class TridaPreTrainedModel(PreTrainedModel):
    config_class = TridaConfig
    base_model_prefix = "model"
    supports_gradient_checkpointing = True
    _no_split_modules = ["TridaDecoderLayer"]
    _skip_keys_device_placement = ["past_key_values"]
    _supports_flash_attn_2 = True
    _supports_sdpa = True
    _supports_flex_attn = True
    _supports_cache_class = True
    _supports_quantized_cache = True
    _supports_static_cache = True
    _supports_attention_backend = True
    _can_record_outputs = {
        "hidden_states": TridaDecoderLayer,
        "attentions": TridaAttention,
    }

    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, TridaRMSNorm):
            module.weight.data.fill_(1.0)


class TridaRotaryEmbedding(nn.Module):
    def __init__(self, config: TridaConfig, device=None):
        super().__init__()
        # BC: "rope_type" was originally "type"
        if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):
            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
        else:
            self.rope_type = "default"
        self.max_seq_len_cached = config.max_position_embeddings
        self.original_max_seq_len = config.max_position_embeddings

        self.config = config
        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]

        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
        self.register_buffer("inv_freq", inv_freq, persistent=False)
        self.original_inv_freq = self.inv_freq

    @torch.no_grad()
    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
    def forward(self, x, position_ids):
        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
        position_ids_expanded = position_ids[:, None, :].float()

        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):  # Force float32
            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
            emb = torch.cat((freqs, freqs), dim=-1)
            cos = emb.cos() * self.attention_scaling
            sin = emb.sin() * self.attention_scaling

        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)



class TridaModel(TridaPreTrainedModel):
    def __init__(self, config: TridaConfig):
        super().__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size
        self.bd_size = config.bd_size

        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
        self.layers = nn.ModuleList(
            [TridaDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = TridaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = TridaRotaryEmbedding(config=config)
        self.gradient_checkpointing = False

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

    def get_input_embeddings(self):
        return self.embed_tokens

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


    def eval_mask(self, seqlen, block_size, cache_seq_len, input_ids=None, mask_id=128012,
                  prev_eval_mask=None):
        """
        Creates attention mask for inference with:
        1. Block-wise causal attention: tokens in current block attend to previous clean blocks
        2. Within-block any-order causal attention: once a token is unmasked, its attention mask is fixed

        Args:
            seqlen: Current sequence length
            block_size: Size of attention blocks
            cache_seq_len: Length of cached sequence from previous blocks
            input_ids: Current input token IDs [batch, seqlen]
            mask_id: Token ID used for masked positions
            prev_eval_mask: Previous attention mask to preserve frozen attention patterns
        """
        device = input_ids.device if input_ids is not None else 'cuda'
        total_len = seqlen + cache_seq_len

        q_indices = torch.arange(seqlen, device=device) + cache_seq_len
        k_indices = torch.arange(total_len, device=device)

        # Block-level causal mask: [seqlen, total_len]
        block_q = q_indices[:, None] // block_size
        block_kv = k_indices[None, :] // block_size
        block_causal_mask = (block_q >= block_kv).unsqueeze(0)  # [1, seqlen, total_len]

        if input_ids is None or mask_id is None:
            return block_causal_mask

        batch_size = input_ids.shape[0]

        # Current mask state
        is_mask_q = (input_ids == mask_id)  # [batch, seqlen]
        is_mask_kv_cached = torch.zeros(batch_size, cache_seq_len, dtype=torch.bool, device=device)
        is_mask_kv = torch.cat([is_mask_kv_cached, is_mask_q], dim=1)  # [batch, total_len]

        # Diagonal (self-attention always allowed)
        diagonal_mask = (q_indices[:, None] == k_indices[None, :]).unsqueeze(0)  # [1, seqlen, total_len]

        if prev_eval_mask is not None:
            # Any-order causal attention:
            # - Currently unmasked tokens: use their frozen attention from prev_eval_mask
            # - Still masked tokens: attend to all currently unmasked tokens + themselves

            is_unmasked_kv = ~is_mask_kv  # [batch, total_len]

            # For still masked queries: attend to all currently unmasked tokens + themselves
            still_masked_q_mask = (block_causal_mask & is_unmasked_kv.unsqueeze(1)) | diagonal_mask

            # Get previous mask (squeeze the head dimension)
            prev_eval_mask_squeezed = prev_eval_mask.squeeze(1)  # [batch, seqlen, total_len]

            # Expand is_mask_q for indexing: [batch, seqlen, total_len]
            is_mask_q_exp = is_mask_q.unsqueeze(-1).expand(-1, -1, total_len)

            # Build final mask:
            # - Currently unmasked tokens: use their frozen attention from prev_eval_mask
            # - Still masked tokens: use still_masked_q_mask
            final_mask = torch.where(
                is_mask_q_exp,
                still_masked_q_mask,
                prev_eval_mask_squeezed
            )

            return final_mask.unsqueeze(1)  # [batch, 1, seqlen, total_len]
        else:
            # First step or no previous mask info
            # Masked tokens: attend to all unmasked tokens + themselves (to predict what to fill in)
            # Unmasked tokens: block causal to other unmasked tokens + self

            is_mask_q_exp = is_mask_q.unsqueeze(-1)  # [batch, seqlen, 1]
            is_mask_kv_exp = is_mask_kv.unsqueeze(1)  # [batch, 1, total_len]

            # Both masked and unmasked queries can see: unmasked keys within block causal + diagonal
            final_mask = (block_causal_mask & ~is_mask_kv_exp) | diagonal_mask

            return final_mask.unsqueeze(1)  # [batch, 1, seqlen, total_len]

    def gen_mask(self, seqlen, block_size, B, H):
        mask = create_block_mask(
            partial(block_diff_mask, block_size=block_size, n=seqlen),
            B=B, H=H, Q_LEN=seqlen*2, KV_LEN=seqlen*2)

        return mask

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        update_past_key_values: Optional[bool] = False,
        block_size: Optional[int] = 32,
        use_block_cache: Optional[bool] = False,
        block_past_key_values: Optional[Cache] = None,
        replace_position: Optional[int] = None,
        prev_attention_mask: Optional[torch.Tensor] = None,
        mask_id: Optional[int] = 128012,
        **kwargs
    ) -> BaseModelOutputWithPast:
        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 use_cache and past_key_values is None:
            past_key_values = DynamicCache()

        if use_block_cache and block_past_key_values is None:
            block_past_key_values = DynamicCache()

        if cache_position is None:
            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
            if self.training:
                cache_position = torch.arange(
                    past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1]//2, device=inputs_embeds.device
                )
            else:
                if use_block_cache:
                    block_start_position = past_seen_tokens+replace_position if replace_position is not None else past_seen_tokens
                    cache_position = torch.arange(
                        block_start_position, block_start_position + inputs_embeds.shape[1], device=inputs_embeds.device
                    )
                else:
                    cache_position = torch.arange(
                        past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1] if not self.training else inputs_embeds.shape[1]//2, device=inputs_embeds.device
                    )

        if position_ids is None:
            position_ids = cache_position.unsqueeze(0)
        
        if self.training:
            attention_mask = self.gen_mask(labels.shape[1], self.bd_size, labels.shape[0], self.config.num_attention_heads).to(device=inputs_embeds.device)
        else:
            # Always compute attention mask for proper any-order causal attention
            cache_seq_len = past_key_values.get_seq_length() if past_key_values is not None else 0
            attention_mask = self.eval_mask(
                input_ids.shape[1],
                block_size,
                cache_seq_len,
                input_ids=input_ids,
                mask_id=mask_id,
                prev_eval_mask=prev_attention_mask
            ).to(device=inputs_embeds.device)

        hidden_states = inputs_embeds

        # create position embeddings to be shared across the decoder layers
        position_embeddings = self.rotary_emb(hidden_states, position_ids)

        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
            hidden_states = decoder_layer(
                hidden_states,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_value=past_key_values,
                use_cache=use_cache,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
                update_past_key_values=update_past_key_values,
                use_block_cache=use_block_cache,
                block_past_key_values=block_past_key_values,
                replace_position=replace_position,
                **kwargs,
            )

        hidden_states = self.norm(hidden_states)
        return BaseModelOutputWithPastAndBlockCache(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values if use_cache else None,
            block_past_key_values=block_past_key_values if use_block_cache else None,
            computed_attention_mask=attention_mask if not self.training else None,
        )


class TridaForDLM(TridaPreTrainedModel, GenerationMixin):
    _tied_weights_keys = ["lm_head.weight"]
    _tp_plan = {"lm_head": "colwise_rep"}
    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}

    def __init__(self, config):
        super().__init__(config)
        self.model = TridaModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

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

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

    def set_input_embeddings(self, value):
        self.model.embed_tokens = 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 _create_attention_mask(self, input_ids, seq_len, pad_token_id):
        """
        Create attention mask that marks padding positions as 0 (masked)
        and real token positions as 1 (unmasked).

        Args:
            input_ids: Tensor of shape (batch_size, seq_length)
            seq_len: Tensor of shape (batch_size,) with actual sequence lengths for each sample
            pad_token_id: Token ID used for padding

        Returns:
            attention_mask: Tensor of shape (batch_size, seq_length) with 1 for real tokens, 0 for padding
        """
        batch_size, seq_length = input_ids.shape
        attention_mask = torch.ones((batch_size, seq_length), dtype=torch.long, device=input_ids.device)

        # Mark padding positions as 0
        for i in range(batch_size):
            # Padding can occur after the actual sequence length
            if seq_len[i] < seq_length:
                attention_mask[i, seq_len[i]:] = 0

        return attention_mask

    @can_return_tuple
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        logits_to_keep: Union[int, torch.Tensor] = 0,
        update_past_key_values: Optional[bool] = False,
        block_size: Optional[int] = 32,
        use_block_cache: Optional[bool] = False,
        block_past_key_values: Optional[Cache] = None,
        replace_position: Optional[int] = None,
        mask_id: Optional[int] = 128012,
        prev_attention_mask: Optional[torch.Tensor] = None,
        **kwargs
    ) -> CausalLMOutputWithPastAndBlockCache:

        if self.training:
            original_labels = labels.clone()
            original_input_ids = input_ids.clone()

            noisy_input_ids = input_ids.clone()

            input_ids = input_ids.reshape(input_ids.shape[0] * input_ids.shape[1] // self.model.bd_size, self.model.bd_size)
            b, l = input_ids.shape
            t = torch.rand((b,), device=input_ids.device)
            eps=1e-3
            p_mask = (1 - eps) * t + eps
            p_mask = p_mask[:, None].repeat(1, l)

            mask_indices = torch.rand((b, l), device=input_ids.device) < p_mask
            x_t = torch.where(mask_indices, mask_id, input_ids).reshape(labels.shape)
            noisy_input_ids[labels != -100] = x_t[labels != -100]
            mask = (noisy_input_ids != mask_id)
            labels[mask] = -100
            input_ids = torch.cat([noisy_input_ids, input_ids.reshape(labels.shape)], dim=1)

            complementary_noisy_input_ids = original_input_ids.clone()
            complementary_labels = original_labels.clone()

            complementary_input_ids = original_input_ids.reshape(original_input_ids.shape[0] * original_input_ids.shape[1] // self.model.bd_size, self.model.bd_size)

            complementary_mask_indices = ~mask_indices
            complementary_x_t = torch.where(complementary_mask_indices, mask_id, complementary_input_ids).reshape(labels.shape)
            complementary_noisy_input_ids[complementary_labels != -100] = complementary_x_t[complementary_labels != -100]
            complementary_mask = (complementary_noisy_input_ids != mask_id)
            complementary_labels[complementary_mask] = -100
            complementary_input_ids = torch.cat([complementary_noisy_input_ids, complementary_input_ids.reshape(complementary_labels.shape)], dim=1)

            input_ids = torch.cat([input_ids, complementary_input_ids], dim=0)
            labels = torch.cat([labels, complementary_labels], dim=0)

        outputs: BaseModelOutputWithPastAndBlockCache = self.model(
            input_ids=input_ids,
            labels=labels,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            cache_position=cache_position,
            update_past_key_values=update_past_key_values,
            block_size=block_size,
            use_block_cache=use_block_cache,
            block_past_key_values=block_past_key_values,
            replace_position=replace_position,
            prev_attention_mask=prev_attention_mask,
            mask_id=mask_id,
            **kwargs,
        )

        hidden_states = outputs.last_hidden_state
        if self.training:
            hidden_states = hidden_states[:, :hidden_states.shape[1]//2, :]
        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
        logits = self.lm_head(hidden_states[:, slice_indices, :])

        loss = None
        if labels is not None:
            # Standard causal LM loss with automatic label shifting (logits[i] predicts labels[i+1])
            # This is correct even for block diffusion training
            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)

        return CausalLMOutputWithPastAndBlockCache(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
            block_past_key_values=outputs.block_past_key_values,
            computed_attention_mask=outputs.computed_attention_mask,
        )

    def _has_stop_token(self, tokens, stop_token):
        """Check if any batch has the stop token. Returns (any_has_stop, per_batch_mask)."""
        has_stop = (tokens == stop_token).any(dim=-1)  # [batch]
        return has_stop.any().item(), has_stop

    def _get_first_stop_token_idx(self, tokens, stop_token):
        """Get the index of the first stop token for each batch. Returns -1 if not found."""
        batch_size, seq_len = tokens.shape
        # Find positions where stop_token occurs
        is_stop = (tokens == stop_token)  # [batch, seq]
        # For each batch, find the first occurrence (or seq_len if not found)
        # Use argmax on the boolean tensor - it returns first True index, or 0 if all False
        has_stop = is_stop.any(dim=-1)  # [batch]
        first_idx = is_stop.float().argmax(dim=-1)  # [batch]
        # Set to -1 for batches without stop token
        first_idx = torch.where(has_stop, first_idx, torch.tensor(-1, device=tokens.device))
        return first_idx  # [batch]

    def _get_per_sequence_stop_status(self, tokens, prompt_length, stop_token, mask_id):
        """
        Check stop condition per sequence.
        Returns a boolean tensor [batch] where True means the sequence should stop.
        A sequence stops when it has a stop token with no mask tokens before it.
        """
        batch_size = tokens.shape[0]
        generated = tokens[:, prompt_length:]

        # Get stop token positions
        first_stop_idx = self._get_first_stop_token_idx(generated, stop_token)  # [batch]
        has_stop = first_stop_idx >= 0  # [batch]

        # For each sequence, check if there are mask tokens before the stop token
        should_stop = torch.zeros(batch_size, dtype=torch.bool, device=tokens.device)
        for b in range(batch_size):
            if has_stop[b]:
                # Check if any mask tokens exist before the stop token
                if (generated[b, :first_stop_idx[b]] == mask_id).sum() == 0:
                    should_stop[b] = True

        return should_stop

    @torch.no_grad()
    def generate(
        self,
        input_ids,
        max_new_tokens,
        mask_id=128012,
        threshold=1,
        small_block_size=8,
        block_size=32,
        stop_token=128001,
        stopping_criteria=None,
        top_p=0.95,
        temperature=0,
        use_block_cache=False,
        pad_token_id=128004,
        return_dict_in_generate=False,
        **kwargs
    ):
        batch_size = input_ids.shape[0]
        num_blocks = max_new_tokens // block_size
        original_input_length = input_ids.shape[1]

        # Track which sequences are finished (have generated stop token with no masks before it)
        finished = torch.zeros(batch_size, dtype=torch.bool, device=self.device)

        if input_ids.shape[1] > block_size:
            output = self.forward(input_ids=input_ids[:, :(input_ids.shape[1] // block_size * block_size)], use_cache=True, update_past_key_values=True, block_size=block_size)
            logits, past_key_values = output.logits, output.past_key_values
            if input_ids.shape[1] % block_size == 0:
                next_token = logits[:, -1:, :].argmax(dim=-1)
                input_ids = torch.cat([input_ids, next_token], dim=1)
        else:
            past_key_values = None

        num_small_blocks = block_size // small_block_size

        for block_idx in range(num_blocks):
            # Check if all sequences are finished
            if finished.all():
                break

            prompt_length = input_ids.shape[1]
            # Initialize x_init with mask_id
            x_init = mask_id * torch.ones((batch_size, block_size - prompt_length % block_size), device=self.device, dtype=torch.long)
            x_init = torch.cat([input_ids, x_init], dim=1)

            x_t = x_init.clone()
            block_past_key_values = None
            block_attention_mask = None  # Track attention mask within block for any-order causal

            while True:
                # Update finished status per sequence
                newly_finished = self._get_per_sequence_stop_status(x_t, prompt_length, stop_token, mask_id)
                finished = finished | newly_finished

                # Check if all sequences are finished
                if finished.all():
                    break

                # Only consider mask positions for unfinished sequences
                mask_idx = (x_t[:, -block_size:] == mask_id)
                # Zero out mask_idx for finished sequences (they don't need more unmasking)
                mask_idx = mask_idx & (~finished.unsqueeze(-1))

                # Decode a complete block, update cache, and generate the next token
                if mask_idx.sum() == 0:
                    output = self.forward(input_ids=x_t[:, -block_size:], use_cache=True, past_key_values=past_key_values, update_past_key_values=True, block_size=block_size)
                    logits, past_key_values = output.logits, output.past_key_values
                    next_token = logits[:, -1:, :].argmax(dim=-1)
                    # Only append next token for unfinished sequences; use pad for finished
                    next_token = torch.where(finished.unsqueeze(-1), pad_token_id, next_token)
                    x_t = torch.cat([x_t, next_token], dim=1)
                    break

                for small_block_idx in range(num_small_blocks):
                    small_block_start_idx = small_block_idx * small_block_size
                    small_block_end_idx = small_block_start_idx + small_block_size

                    start = -block_size + small_block_start_idx
                    end = None if block_size == small_block_end_idx else -block_size + small_block_end_idx

                    while True:
                        # Recompute mask_idx considering finished sequences
                        mask_idx = (x_t[:, -block_size:] == mask_id)
                        mask_idx = mask_idx & (~finished.unsqueeze(-1))

                        if mask_idx[:, start:end].sum() == 0:
                            break

                        # Update finished status
                        newly_finished = self._get_per_sequence_stop_status(x_t, prompt_length, stop_token, mask_id)
                        finished = finished | newly_finished

                        if finished.all():
                            break

                        if use_block_cache:
                            if block_past_key_values is None or (x_t[:, -block_size+small_block_start_idx] == mask_id).any():
                                output = self.forward(input_ids=x_t[:, -block_size:], use_cache=True, past_key_values=past_key_values, update_past_key_values=False, use_block_cache=True, prev_attention_mask=block_attention_mask, mask_id=mask_id)
                                logits, block_past_key_values = output.logits, output.block_past_key_values
                                block_attention_mask = output.computed_attention_mask  # Track attention mask for any-order causal
                                logits = torch.cat([logits[:, :1, :], logits[:, :-1, :]], dim=1)
                                logits = logits[:, start:end]
                            else:
                                output = self.forward(input_ids=x_t[:,start:end], use_cache=True, past_key_values=past_key_values, update_past_key_values=False, use_block_cache=True, block_past_key_values=block_past_key_values, replace_position=small_block_start_idx, prev_attention_mask=block_attention_mask, mask_id=mask_id)
                                logits = output.logits
                                block_attention_mask = output.computed_attention_mask  # Track attention mask for any-order causal
                                logits = torch.cat([logits[:, :1, :], logits[:, :-1, :]], dim=1)
                        else:
                            output = self.forward(input_ids=x_t[:, -block_size:], use_cache=True, past_key_values=past_key_values, update_past_key_values=False, prev_attention_mask=block_attention_mask, mask_id=mask_id)
                            logits = output.logits
                            block_attention_mask = output.computed_attention_mask  # Track attention mask for any-order causal
                            logits = torch.cat([logits[:, :1, :], logits[:, :-1, :]], dim=1)
                            logits = logits[:, start:end]

                        # Allow stop_token to be generated naturally (don't block it)
                        x_1, p_1t = self.sample_with_top_p(logits, top_p=top_p, temperature=temperature, stop_token_id=None)
                        # Select tokens with probability greater than threshold from p_1t
                        x1_p = torch.squeeze(torch.gather(p_1t, dim=-1, index=torch.unsqueeze(x_1, -1)), -1)
                        x1_p = torch.where(mask_idx[:, start:end], x1_p, -torch.inf)

                        unmask_idx = (x1_p > threshold)
                        max_prob_idx = x1_p.argmax(dim=-1)
                        unmask_idx[torch.arange(x_1.shape[0], device=self.device), max_prob_idx] = True
                        unmask_idx = unmask_idx & mask_idx[:, start:end]

                        # Don't update tokens for finished sequences
                        unmask_idx = unmask_idx & (~finished.unsqueeze(-1))

                        x_t[:, start:end][unmask_idx] = x_1[unmask_idx]

            input_ids = x_t

        # Per-sequence truncation: truncate each sequence at its stop token
        # Pad shorter sequences to maintain consistent tensor shape
        first_stop_idx = self._get_first_stop_token_idx(input_ids[:, original_input_length:], stop_token)

        # Calculate the max output length (considering sequences without stop token)
        max_output_len = input_ids.shape[1]
        has_stop = first_stop_idx >= 0
        if has_stop.any():
            # For sequences with stop token: original_input_length + stop_idx + 1 (include stop token)
            # For sequences without: keep full length
            output_lens = torch.where(
                has_stop,
                original_input_length + first_stop_idx + 1,
                torch.tensor(input_ids.shape[1], device=self.device)
            )
            max_output_len = output_lens.max().item()

        # Truncate to max_output_len
        input_ids = input_ids[:, :max_output_len]

        # Replace tokens after each sequence's stop token with pad_token_id
        for b in range(batch_size):
            if has_stop[b]:
                stop_pos = original_input_length + first_stop_idx[b].item() + 1
                if stop_pos < input_ids.shape[1]:
                    input_ids[b, stop_pos:] = pad_token_id

        return input_ids

    def sample_with_top_p(self, logits, top_p=0.95, temperature=1.0, mask_token_id=128012, pad_token_id=128004, stop_token_id=None):
        # Mask out special tokens to prevent them from being generated during unmasking
        # mask_token and pad_token should never be generated as actual output
        # stop_token_id can be optionally masked during intra-block unmasking to prevent premature stopping
        logits = logits.clone()
        logits[..., mask_token_id] = -float('inf')
        logits[..., pad_token_id] = -float('inf')
        if stop_token_id is not None:
            logits[..., stop_token_id] = -float('inf')

        # Calculate probabilities
        if temperature > 0:
            scaled_logits = logits / temperature
        else:
            p_1t = torch.softmax(logits, dim=-1)
            x_1 = p_1t.argmax(dim=-1)
            return x_1, p_1t

        probs = F.softmax(scaled_logits, dim=-1)

        sorted_probs, sorted_indices = torch.sort(probs, descending=True)
        cumulative_probs = torch.cumsum(sorted_probs, dim=-1)

        sorted_indices_to_remove = cumulative_probs > top_p
        sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
        sorted_indices_to_remove[..., 0] = 0

        indices_to_remove = torch.zeros_like(probs, dtype=torch.bool).scatter_(
            dim=-1, index=sorted_indices, src=sorted_indices_to_remove
        )
        
        probs[indices_to_remove] = 0

        # Renormalize so that the probabilities of remaining tokens sum to 1
        # Add a small epsilon value to prevent division by zero
        probs_sum = torch.sum(probs, dim=-1, keepdim=True)
        probs_sum = torch.clamp(probs_sum, min=1e-10)
        normalized_probs = probs / probs_sum

        p_1t = normalized_probs
        # Handle arbitrary batch sizes: reshape to [batch * seq, vocab], sample, reshape back
        batch_size, seq_len, vocab_size = p_1t.shape
        p_1t_flat = p_1t.view(-1, vocab_size)
        x_1 = torch.multinomial(p_1t_flat, num_samples=1).view(batch_size, seq_len)

        return x_1, p_1t