import itertools
import logging
from collections.abc import Iterable
from types import SimpleNamespace

import torch
import torch.nn as nn
import torch.nn.functional as F
import vllm._custom_ops as ops
from diffusers.models.activations import get_activation
from diffusers.models.embeddings import Timesteps, get_1d_rotary_pos_embed
from diffusers.models.modeling_outputs import Transformer2DModelOutput
from einops import rearrange, repeat
from vllm.model_executor.layers.activation import get_act_and_mul_fn
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
    MergedColumnParallelLinear,
    QKVParallelLinear,
    RowParallelLinear,
)
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.model_loader.weight_utils import default_weight_loader

from vllm_omni.diffusion.attention.layer import Attention
from vllm_omni.diffusion.distributed.sp_plan import (
    SequenceParallelInput,
    SequenceParallelOutput,
)
from vllm_omni.platforms import current_omni_platform

logger = logging.getLogger(__name__)


def _patch_cutlass_padded_fp8():
    """Monkey-patch vllm._custom_ops.cutlass_scaled_mm to pad tensors whose
    dimensions are not multiples of 16, so the CUTLASS FP8 kernel is used.

    OmniGen2 has hidden_size=2520 (2520 % 16 == 8).  Without this patch,
    vLLM's cutlass_scaled_mm falls back to a Triton scaled_mm kernel for
    every FP8 linear layer (QKV, attn output, gate_up_proj, down_proj),
    which is dramatically slower than the native CUTLASS FP8 tensor-core
    path on H100/H200 GPUs.

    Weight tensors (b) are constant across forward passes, so padded
    versions are computed once and cached by data_ptr to avoid repeated
    allocation and column-major conversion overhead.
    """
    _orig_cutlass_scaled_mm = ops.cutlass_scaled_mm
    # Cache: data_ptr → (padded_b, padded_bias, padded_scale_b, pad_k, pad_n, orig_n)
    _weight_cache: dict[int, tuple] = {}

    def _padded_cutlass_scaled_mm(
        a: torch.Tensor,
        b: torch.Tensor,
        scale_a: torch.Tensor,
        scale_b: torch.Tensor,
        out_dtype: torch.dtype,
        bias: torch.Tensor | None = None,
    ) -> torch.Tensor:
        if b.shape[0] % 16 == 0 and b.shape[1] % 16 == 0:
            return _orig_cutlass_scaled_mm(a, b, scale_a, scale_b, out_dtype, bias)

        # Reshape to 2D (mirrors the original function)
        target_shape = (*a.shape[:-1], b.shape[1])
        a = a.view(-1, a.shape[-1])
        orig_n = b.shape[1]

        # Cache the padded weight — it's a model parameter that never changes.
        key = b.data_ptr()
        if key not in _weight_cache:
            pad_k = (16 - b.shape[0] % 16) % 16
            pad_n = (16 - orig_n % 16) % 16
            b_pad = b
            if pad_k > 0:
                b_pad = F.pad(b_pad, (0, 0, 0, pad_k))
            if pad_n > 0:
                b_pad = F.pad(b_pad, (0, pad_n))
            # CUTLASS requires b column-major (stride(0)==1).
            b_pad = b_pad.t().contiguous().t()

            bias_pad = None
            if bias is not None and pad_n > 0:
                bias_pad = F.pad(bias, (0, pad_n))

            scale_b_pad = scale_b
            if scale_b.numel() > 1 and pad_n > 0:
                scale_b_pad = F.pad(
                    scale_b.view(-1, scale_b.shape[-1]),
                    (0, pad_n),
                    value=1.0,
                )

            _weight_cache[key] = (
                b_pad,
                bias_pad,
                scale_b_pad,
                pad_k,
                pad_n,
                orig_n,
            )

        b_pad, bias_pad, scale_b_pad, pad_k, pad_n, orig_n = _weight_cache[key]

        # Pad activations on K dimension (cheap — activations are small).
        if pad_k > 0:
            a = F.pad(a, (0, pad_k)).contiguous()

        out = torch.empty((a.shape[0], b_pad.shape[1]), dtype=out_dtype, device=a.device)
        torch.ops._C.cutlass_scaled_mm(
            out,
            a,
            b_pad,
            scale_a,
            scale_b_pad,
            bias_pad if bias is not None else None,
        )

        if pad_n > 0:
            out = out[:, :orig_n]

        return out.view(*target_shape)

    ops.cutlass_scaled_mm = _padded_cutlass_scaled_mm
    logger.info(
        "Patched vllm._custom_ops.cutlass_scaled_mm with CUTLASS-padded FP8 "
        "variant (avoids slow Triton fallback for non-%%16 dimensions)"
    )


_patch_cutlass_padded_fp8()


class OmniGen2Attention(nn.Module):
    def __init__(
        self,
        dim: int,
        num_heads: int,
        num_kv_heads: int,
        eps: float = 1e-5,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ):
        super().__init__()
        self.dim = dim
        self.num_heads = num_heads
        self.num_kv_heads = num_kv_heads
        self.head_dim = dim // num_heads
        self.eps = eps

        self.to_qkv = QKVParallelLinear(
            hidden_size=dim,
            head_size=self.head_dim,
            total_num_heads=num_heads,
            total_num_kv_heads=num_kv_heads,
            disable_tp=True,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.to_qkv",
        )

        self.norm_q = RMSNorm(self.head_dim, eps=eps)
        self.norm_k = RMSNorm(self.head_dim, eps=eps)

        self.to_out = nn.ModuleList(
            [
                RowParallelLinear(
                    dim,
                    dim,
                    bias=False,
                    input_is_parallel=False,
                    quant_config=quant_config,
                    return_bias=False,
                    prefix=f"{prefix}.to_out.0",
                )
            ]
        )
        self.attn = Attention(
            num_heads=num_heads,
            head_size=self.head_dim,
            softmax_scale=1.0 / (self.head_dim**0.5),
            causal=False,
        )

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        image_rotary_emb: torch.Tensor | None = None,
    ) -> torch.Tensor:
        """
        Process attention computation with flash attention.

        Args:
            hidden_states: Hidden states tensor of shape (batch_size, seq_len, hidden_dim)
            attention_mask: Optional attention mask tensor
            image_rotary_emb: Optional rotary embeddings for image tokens

        Returns:
            torch.Tensor: Processed hidden states after attention computation
        """
        batch_size = hidden_states.shape[0]

        # Contiguous layout for FP8 quantized linear GEMMs (matches FLUX DiT).
        hidden_states = hidden_states.contiguous()

        # Get Query-Key-Value Pair
        qkv, _ = self.to_qkv(hidden_states)

        q_size = self.num_heads * self.head_dim
        kv_size = self.num_kv_heads * self.head_dim
        query, key, value = qkv.split([q_size, kv_size, kv_size], dim=-1)

        # Reshape tensors for attention computation
        query = query.unflatten(-1, (self.num_heads, -1))
        key = key.unflatten(-1, (self.num_kv_heads, -1))
        value = value.unflatten(-1, (self.num_kv_heads, -1))

        # Apply Query-Key normalization
        if self.norm_q is not None:
            query = self.norm_q(query)
        if self.norm_k is not None:
            key = self.norm_k(key)

        dtype = query.dtype

        # Apply Rotary Position Embeddings
        if image_rotary_emb is not None:
            query = apply_rotary_emb(query, image_rotary_emb, use_real=False)
            key = apply_rotary_emb(key, image_rotary_emb, use_real=False)

        query, key = query.to(dtype), key.to(dtype)

        if self.num_kv_heads < self.num_heads:
            # Repeat key and value heads to match query heads
            num_groups = self.num_heads // self.num_kv_heads
            key = key.repeat_interleave(num_groups, dim=2)
            value = value.repeat_interleave(num_groups, dim=2)

        hidden_states = self.attn(
            query,
            key,
            value,
        )

        # Reshape back
        hidden_states = hidden_states.reshape(batch_size, -1, self.num_heads * self.head_dim)
        hidden_states = hidden_states.to(dtype)

        hidden_states = self.to_out[0](hidden_states.contiguous())

        return hidden_states


class TimestepEmbedding(nn.Module):
    def __init__(
        self,
        in_channels: int,
        time_embed_dim: int,
        act_fn: str = "silu",
        out_dim: int = None,
        post_act_fn: str | None = None,
        cond_proj_dim=None,
        sample_proj_bias=True,
    ):
        super().__init__()

        self.linear_1 = nn.Linear(in_channels, time_embed_dim, sample_proj_bias)

        if cond_proj_dim is not None:
            self.cond_proj = nn.Linear(cond_proj_dim, in_channels, bias=False)
        else:
            self.cond_proj = None

        self.act = get_activation(act_fn)

        if out_dim is not None:
            time_embed_dim_out = out_dim
        else:
            time_embed_dim_out = time_embed_dim
        self.linear_2 = nn.Linear(time_embed_dim, time_embed_dim_out, sample_proj_bias)

        if post_act_fn is None:
            self.post_act = None
        else:
            self.post_act = get_activation(post_act_fn)

    def forward(self, sample, condition=None):
        if condition is not None:
            sample = sample + self.cond_proj(condition)
        sample = self.linear_1(sample)
        sample = self.act(sample)
        sample = self.linear_2(sample)

        if self.post_act is not None:
            sample = self.post_act(sample)
        return sample


def apply_rotary_emb(
    x: torch.Tensor,
    freqs_cis: torch.Tensor | tuple[torch.Tensor],
    use_real: bool = True,
    use_real_unbind_dim: int = -1,
) -> tuple[torch.Tensor, torch.Tensor]:
    """
    Apply rotary embeddings to input tensors using the given frequency tensor. This function applies rotary embeddings
    to the given query or key 'x' tensors using the provided frequency tensor 'freqs_cis'. The input tensors are
    reshaped as complex numbers, and the frequency tensor is reshaped for broadcasting compatibility. The resulting
    tensors contain rotary embeddings and are returned as real tensors.

    Args:
        x (`torch.Tensor`):
            Query or key tensor to apply rotary embeddings. [B, H, S, D] xk (torch.Tensor): Key tensor to apply
        freqs_cis (`Tuple[torch.Tensor]`): Precomputed frequency tensor for complex exponentials. ([S, D], [S, D],)

    Returns:
        Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor and key tensor with rotary embeddings.
    """
    if use_real:
        cos, sin = freqs_cis  # [S, D]
        cos = cos[None, None]
        sin = sin[None, None]
        cos, sin = cos.to(x.device), sin.to(x.device)

        if use_real_unbind_dim == -1:
            # Used for flux, cogvideox, hunyuan-dit
            x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)  # [B, S, H, D//2]
            x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
        elif use_real_unbind_dim == -2:
            # Used for Stable Audio, OmniGen and CogView4
            x_real, x_imag = x.reshape(*x.shape[:-1], 2, -1).unbind(-2)  # [B, S, H, D//2]
            x_rotated = torch.cat([-x_imag, x_real], dim=-1)
        else:
            raise ValueError(f"`use_real_unbind_dim={use_real_unbind_dim}` but should be -1 or -2.")

        out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype)

        return out
    else:
        # used for lumina
        x_rotated = torch.view_as_complex(x.float().reshape(*x.shape[:-1], x.shape[-1] // 2, 2))
        freqs_cis = freqs_cis.unsqueeze(2)
        x_out = torch.view_as_real(x_rotated * freqs_cis).flatten(3)

        return x_out.type_as(x)


class LuminaRMSNormZero(nn.Module):
    """
    Norm layer adaptive RMS normalization zero.

    Parameters:
        embedding_dim (`int`): The size of each embedding vector.
    """

    def __init__(
        self,
        embedding_dim: int,
        norm_eps: float,
        norm_elementwise_affine: bool,
        **kwargs,
    ):
        super().__init__()
        self.silu = nn.SiLU()
        self.linear = nn.Linear(
            min(embedding_dim, 1024),
            4 * embedding_dim,
            bias=True,
        )

        self.norm = RMSNorm(embedding_dim, eps=norm_eps)

    def forward(
        self,
        x: torch.Tensor,
        emb: torch.Tensor | None = None,
    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
        emb = self.linear(self.silu(emb))
        scale_msa, gate_msa, scale_mlp, gate_mlp = emb.chunk(4, dim=1)
        x = self.norm(x) * (1 + scale_msa[:, None])
        return x, gate_msa, scale_mlp, gate_mlp


class LuminaLayerNormContinuous(nn.Module):
    def __init__(
        self,
        embedding_dim: int,
        conditioning_embedding_dim: int,
        # NOTE: It is a bit weird that the norm layer can be configured to have scale and shift parameters
        # because the output is immediately scaled and shifted by the projected conditioning embeddings.
        # Note that AdaLayerNorm does not let the norm layer have scale and shift parameters.
        # However, this is how it was implemented in the original code, and it's rather likely you should
        # set `elementwise_affine` to False.
        elementwise_affine=True,
        eps=1e-5,
        bias=True,
        norm_type="layer_norm",
        out_dim: int | None = None,
    ):
        super().__init__()

        # AdaLN
        self.silu = nn.SiLU()
        self.linear_1 = nn.Linear(conditioning_embedding_dim, embedding_dim, bias=bias)

        if norm_type == "layer_norm":
            self.norm = nn.LayerNorm(embedding_dim, eps, elementwise_affine, bias)
        elif norm_type == "rms_norm":
            self.norm = RMSNorm(embedding_dim, eps=eps)
        else:
            raise ValueError(f"unknown norm_type {norm_type}")

        self.linear_2 = None
        if out_dim is not None:
            self.linear_2 = nn.Linear(embedding_dim, out_dim, bias=bias)

    def forward(
        self,
        x: torch.Tensor,
        conditioning_embedding: torch.Tensor,
    ) -> torch.Tensor:
        # convert back to the original dtype in case `conditioning_embedding`` is upcasted to float32
        # (needed for hunyuanDiT)
        emb = self.linear_1(self.silu(conditioning_embedding).to(x.dtype))
        scale = emb
        x = self.norm(x) * (1 + scale)[:, None, :]

        if self.linear_2 is not None:
            x = self.linear_2(x)

        return x


class LuminaFeedForward(nn.Module):
    r"""
    A feed-forward layer.

    Parameters:
        dim (`int`):
            The dimensionality of the input and output tensors.
        inner_dim (`int`): The intermediate dimension of the feedforward layer.
        multiple_of (`int`, *optional*): Value to ensure hidden dimension is a multiple
            of this value.
        ffn_dim_multiplier (float, *optional*): Custom multiplier for hidden
            dimension. Defaults to None.
    """

    def __init__(
        self,
        dim: int,
        inner_dim: int,
        multiple_of: int | None = 256,
        ffn_dim_multiplier: float | None = None,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ):
        super().__init__()

        # custom hidden_size factor multiplier
        if ffn_dim_multiplier is not None:
            inner_dim = int(ffn_dim_multiplier * inner_dim)
        inner_dim = multiple_of * ((inner_dim + multiple_of - 1) // multiple_of)

        self.gate_up_proj = MergedColumnParallelLinear(
            dim,
            [inner_dim, inner_dim],
            bias=False,
            return_bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj",
        )
        self.act_fn = get_act_and_mul_fn("silu")
        self.down_proj = RowParallelLinear(
            inner_dim,
            dim,
            bias=False,
            input_is_parallel=True,
            return_bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.down_proj",
        )

    def forward(self, x):
        x = self.gate_up_proj(x)
        x = self.act_fn(x)
        return self.down_proj(x)


class Lumina2CombinedTimestepCaptionEmbedding(nn.Module):
    def __init__(
        self,
        hidden_size: int = 4096,
        text_feat_dim: int = 2048,
        frequency_embedding_size: int = 256,
        norm_eps: float = 1e-5,
        timestep_scale: float = 1.0,
    ) -> None:
        super().__init__()

        self.time_proj = Timesteps(
            num_channels=frequency_embedding_size,
            flip_sin_to_cos=True,
            downscale_freq_shift=0.0,
            scale=timestep_scale,
        )

        self.timestep_embedder = TimestepEmbedding(
            in_channels=frequency_embedding_size, time_embed_dim=min(hidden_size, 1024)
        )

        self.caption_embedder = nn.Sequential(
            RMSNorm(text_feat_dim, eps=norm_eps),
            nn.Linear(text_feat_dim, hidden_size, bias=True),
        )

    def forward(
        self,
        timestep: torch.Tensor,
        text_hidden_states: torch.Tensor,
        dtype: torch.dtype,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        timestep_proj = self.time_proj(timestep).to(dtype=dtype)
        time_embed = self.timestep_embedder(timestep_proj)
        caption_embed = self.caption_embedder(text_hidden_states)
        return time_embed, caption_embed


class OmniGen2RotaryPosEmbed(nn.Module):
    def __init__(
        self,
        theta: int,
        axes_dim: tuple[int, int, int],
        axes_lens: tuple[int, int, int] = (300, 512, 512),
        patch_size: int = 2,
    ):
        super().__init__()
        self.theta = theta
        self.axes_dim = axes_dim
        self.axes_lens = axes_lens
        self.patch_size = patch_size

    @staticmethod
    def get_freqs_cis(
        axes_dim: tuple[int, int, int], axes_lens: tuple[int, int, int], theta: int
    ) -> list[torch.Tensor]:
        freqs_cis = []
        freqs_dtype = torch.float64 if current_omni_platform.supports_float64() else torch.float32
        for i, (d, e) in enumerate(zip(axes_dim, axes_lens)):
            emb = get_1d_rotary_pos_embed(d, e, theta=theta, freqs_dtype=freqs_dtype)
            freqs_cis.append(emb)
        return freqs_cis

    def _get_freqs_cis(self, freqs_cis, ids: torch.Tensor) -> torch.Tensor:
        device = ids.device
        if ids.device.type == "mps":
            ids = ids.to("cpu")

        result = []
        for i in range(len(self.axes_dim)):
            freqs = freqs_cis[i].to(ids.device)
            index = ids[:, :, i : i + 1].repeat(1, 1, freqs.shape[-1]).to(torch.int64)
            result.append(torch.gather(freqs.unsqueeze(0).repeat(index.shape[0], 1, 1), dim=1, index=index))
        return torch.cat(result, dim=-1).to(device)

    def forward(
        self,
        freqs_cis,
        attention_mask,
        l_effective_ref_img_len,
        l_effective_img_len,
        ref_img_sizes,
        img_sizes,
        device,
    ):
        batch_size = len(attention_mask)
        p = self.patch_size

        encoder_seq_len = attention_mask.shape[1]
        l_effective_cap_len = attention_mask.sum(dim=1).tolist()

        seq_lengths = [
            cap_len + sum(ref_img_len) + img_len
            for cap_len, ref_img_len, img_len in zip(l_effective_cap_len, l_effective_ref_img_len, l_effective_img_len)
        ]

        max_seq_len = max(seq_lengths)
        max_ref_img_len = max([sum(ref_img_len) for ref_img_len in l_effective_ref_img_len])
        max_img_len = max(l_effective_img_len)

        # Create position IDs
        position_ids = torch.zeros(batch_size, max_seq_len, 3, dtype=torch.int32, device=device)

        for i, (cap_seq_len, seq_len) in enumerate(zip(l_effective_cap_len, seq_lengths)):
            # add text position ids
            position_ids[i, :cap_seq_len] = repeat(
                torch.arange(cap_seq_len, dtype=torch.int32, device=device), "l -> l 3"
            )

            pe_shift = cap_seq_len
            pe_shift_len = cap_seq_len

            if ref_img_sizes[i] is not None:
                for ref_img_size, ref_img_len in zip(ref_img_sizes[i], l_effective_ref_img_len[i]):
                    H, W = ref_img_size
                    ref_H_tokens, ref_W_tokens = H // p, W // p
                    assert ref_H_tokens * ref_W_tokens == ref_img_len
                    # add image position ids

                    row_ids = repeat(
                        torch.arange(ref_H_tokens, dtype=torch.int32, device=device),
                        "h -> h w",
                        w=ref_W_tokens,
                    ).flatten()
                    col_ids = repeat(
                        torch.arange(ref_W_tokens, dtype=torch.int32, device=device),
                        "w -> h w",
                        h=ref_H_tokens,
                    ).flatten()
                    position_ids[i, pe_shift_len : pe_shift_len + ref_img_len, 0] = pe_shift
                    position_ids[i, pe_shift_len : pe_shift_len + ref_img_len, 1] = row_ids
                    position_ids[i, pe_shift_len : pe_shift_len + ref_img_len, 2] = col_ids

                    pe_shift += max(ref_H_tokens, ref_W_tokens)
                    pe_shift_len += ref_img_len

            H, W = img_sizes[i]
            H_tokens, W_tokens = H // p, W // p
            assert H_tokens * W_tokens == l_effective_img_len[i]

            row_ids = repeat(
                torch.arange(H_tokens, dtype=torch.int32, device=device),
                "h -> h w",
                w=W_tokens,
            ).flatten()
            col_ids = repeat(
                torch.arange(W_tokens, dtype=torch.int32, device=device),
                "w -> h w",
                h=H_tokens,
            ).flatten()

            assert pe_shift_len + l_effective_img_len[i] == seq_len
            position_ids[i, pe_shift_len:seq_len, 0] = pe_shift
            position_ids[i, pe_shift_len:seq_len, 1] = row_ids
            position_ids[i, pe_shift_len:seq_len, 2] = col_ids

        # Get combined rotary embeddings
        freqs_cis = self._get_freqs_cis(freqs_cis, position_ids)

        # create separate rotary embeddings for captions and images
        cap_freqs_cis = torch.zeros(
            batch_size,
            encoder_seq_len,
            freqs_cis.shape[-1],
            device=device,
            dtype=freqs_cis.dtype,
        )
        ref_img_freqs_cis = torch.zeros(
            batch_size,
            max_ref_img_len,
            freqs_cis.shape[-1],
            device=device,
            dtype=freqs_cis.dtype,
        )
        img_freqs_cis = torch.zeros(
            batch_size,
            max_img_len,
            freqs_cis.shape[-1],
            device=device,
            dtype=freqs_cis.dtype,
        )

        for i, (cap_seq_len, ref_img_len, img_len, seq_len) in enumerate(
            zip(
                l_effective_cap_len,
                l_effective_ref_img_len,
                l_effective_img_len,
                seq_lengths,
            )
        ):
            cap_freqs_cis[i, :cap_seq_len] = freqs_cis[i, :cap_seq_len]
            ref_img_freqs_cis[i, : sum(ref_img_len)] = freqs_cis[i, cap_seq_len : cap_seq_len + sum(ref_img_len)]
            img_freqs_cis[i, :img_len] = freqs_cis[
                i,
                cap_seq_len + sum(ref_img_len) : cap_seq_len + sum(ref_img_len) + img_len,
            ]

        return (
            cap_freqs_cis,
            ref_img_freqs_cis,
            img_freqs_cis,
            freqs_cis,
            l_effective_cap_len,
            seq_lengths,
        )


class OmniGen2TransformerBlock(nn.Module):
    """
    Transformer block for OmniGen2 model.

    This block implements a transformer layer with:
    - Multi-head attention with flash attention
    - Feed-forward network with SwiGLU activation
    - RMS normalization
    - Optional modulation for conditional generation

    Args:
        dim: Dimension of the input and output tensors
        num_attention_heads: Number of attention heads
        num_kv_heads: Number of key-value heads
        multiple_of: Multiple of which the hidden dimension should be
        ffn_dim_multiplier: Multiplier for the feed-forward network dimension
        norm_eps: Epsilon value for normalization layers
        modulation: Whether to use modulation for conditional generation
    """

    def __init__(
        self,
        dim: int,
        num_attention_heads: int,
        num_kv_heads: int,
        multiple_of: int,
        ffn_dim_multiplier: float,
        norm_eps: float,
        modulation: bool = True,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
    ) -> None:
        """Initialize the transformer block."""
        super().__init__()
        self.head_dim = dim // num_attention_heads
        self.modulation = modulation

        self.attn = OmniGen2Attention(
            dim=dim,
            num_heads=num_attention_heads,
            num_kv_heads=num_kv_heads,
            eps=1e-5,
            quant_config=quant_config,
            prefix=f"{prefix}.attn",
        )

        # Initialize feed-forward network
        self.feed_forward = LuminaFeedForward(
            dim=dim,
            inner_dim=4 * dim,
            multiple_of=multiple_of,
            ffn_dim_multiplier=ffn_dim_multiplier,
            quant_config=quant_config,
            prefix=f"{prefix}.feed_forward",
        )

        # Initialize normalization layers
        if modulation:
            self.norm1 = LuminaRMSNormZero(
                embedding_dim=dim,
                norm_eps=norm_eps,
                norm_elementwise_affine=True,
                quant_config=quant_config,
                prefix=f"{prefix}.norm1",
            )
        else:
            self.norm1 = RMSNorm(dim, eps=norm_eps)

        self.ffn_norm1 = RMSNorm(dim, eps=norm_eps)
        self.norm2 = RMSNorm(dim, eps=norm_eps)
        self.ffn_norm2 = RMSNorm(dim, eps=norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor,
        image_rotary_emb: torch.Tensor,
        temb: torch.Tensor | None = None,
    ) -> torch.Tensor:
        """
        Forward pass of the transformer block.

        Args:
            hidden_states: Input hidden states tensor
            attention_mask: Attention mask tensor
            image_rotary_emb: Rotary embeddings for image tokens
            temb: Optional timestep embedding tensor

        Returns:
            torch.Tensor: Output hidden states after transformer block processing
        """
        if self.modulation:
            if temb is None:
                raise ValueError("temb must be provided when modulation is enabled")

            norm_hidden_states, gate_msa, scale_mlp, gate_mlp = self.norm1(hidden_states, temb)
            attn_output = self.attn(
                hidden_states=norm_hidden_states,
                attention_mask=attention_mask,
                image_rotary_emb=image_rotary_emb,
            )
            hidden_states = hidden_states + gate_msa.unsqueeze(1).tanh() * self.norm2(attn_output)
            mlp_output = self.feed_forward(self.ffn_norm1(hidden_states) * (1 + scale_mlp.unsqueeze(1)))
            hidden_states = hidden_states + gate_mlp.unsqueeze(1).tanh() * self.ffn_norm2(mlp_output)
        else:
            norm_hidden_states = self.norm1(hidden_states)
            attn_output = self.attn(
                hidden_states=norm_hidden_states,
                attention_mask=attention_mask,
                image_rotary_emb=image_rotary_emb,
            )
            hidden_states = hidden_states + self.norm2(attn_output)
            mlp_output = self.feed_forward(self.ffn_norm1(hidden_states))
            hidden_states = hidden_states + self.ffn_norm2(mlp_output)

        return hidden_states


class _OmniGen2SPInputBoundary(nn.Module):
    """SP input-sharding boundary marker for ``OmniGen2Transformer2DModel._sp_plan``.

    Resets ``sp_original_seq_len`` on every call: OmniGen2 runs the transformer
    once per CFG branch and the unconditional / conditional branches have
    different joint sequence lengths (text prefix differs). The SP framework's
    auto-pad records the length only on the first call, so without this reset
    a stale length from an earlier branch truncates the gather of later ones.
    """

    def forward(self, hidden_states: torch.Tensor, rotary_emb: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        from vllm_omni.diffusion.forward_context import (
            get_forward_context,
            is_forward_context_available,
        )

        if is_forward_context_available():
            ctx = get_forward_context()
            ctx.sp_original_seq_len = None
            ctx.sp_padding_size = 0
        return hidden_states, rotary_emb


class _OmniGen2SPOutputBoundary(nn.Module):
    """SP output-gathering boundary marker for ``OmniGen2Transformer2DModel._sp_plan``."""

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        return hidden_states


class OmniGen2Transformer2DModel(nn.Module):
    """
    OmniGen2 Transformer 2D Model.

    A transformer-based diffusion model for image generation with:
    - Patch-based image processing
    - Rotary position embeddings
    - Multi-head attention
    - Conditional generation support

    Args:
        patch_size: Size of image patches
        in_channels: Number of input channels
        out_channels: Number of output channels (defaults to in_channels)
        hidden_size: Size of hidden layers
        num_layers: Number of transformer layers
        num_refiner_layers: Number of refiner layers
        num_attention_heads: Number of attention heads
        num_kv_heads: Number of key-value heads
        multiple_of: Multiple of which the hidden dimension should be
        ffn_dim_multiplier: Multiplier for feed-forward network dimension
        norm_eps: Epsilon value for normalization layers
        axes_dim_rope: Dimensions for rotary position embeddings
        axes_lens: Lengths for rotary position embeddings
        text_feat_dim: Dimension of text features
        timestep_scale: Scale factor for timestep embeddings
    """

    # ``attention_mask`` is intentionally omitted from the plan: OmniGen2Attention
    # accepts but never reads it. ``auto_pad=True`` requires Ulysses SP — Ring SP
    # cannot use attention masks.
    _sp_plan = {
        "sp_input_boundary": {
            0: SequenceParallelInput(split_dim=1, expected_dims=3, split_output=True, auto_pad=True),
            1: SequenceParallelInput(split_dim=1, expected_dims=3, split_output=True, auto_pad=True),
        },
        "sp_output_boundary": SequenceParallelOutput(gather_dim=1, expected_dims=3),
    }

    def __init__(
        self,
        patch_size: int = 2,
        in_channels: int = 16,
        out_channels: int | None = None,
        hidden_size: int = 2520,
        num_layers: int = 32,
        num_refiner_layers: int = 2,
        num_attention_heads: int = 21,
        num_kv_heads: int = 7,
        multiple_of: int = 256,
        ffn_dim_multiplier: float | None = None,
        norm_eps: float = 1e-5,
        axes_dim_rope: tuple[int, int, int] = (40, 40, 40),
        axes_lens: tuple[int, int, int] = (1024, 1664, 1664),
        text_feat_dim: int = 2048,
        timestep_scale: float = 1000.0,
        quant_config: QuantizationConfig | None = None,
    ) -> None:
        """Initialize the OmniGen2 transformer model."""
        super().__init__()

        # Validate configuration
        if (hidden_size // num_attention_heads) != sum(axes_dim_rope):
            raise ValueError(
                f"hidden_size // num_attention_heads ({hidden_size // num_attention_heads}) "
                f"must equal sum(axes_dim_rope) ({sum(axes_dim_rope)})"
            )

        self.config = SimpleNamespace(
            patch_size=patch_size,
            in_channels=in_channels,
            out_channels=out_channels or in_channels,
            hidden_size=hidden_size,
            axes_dim_rope=axes_dim_rope,
            axes_lens=axes_lens,
        )
        self.out_channels = self.config.out_channels

        # Initialize embeddings
        self.rope_embedder = OmniGen2RotaryPosEmbed(
            theta=10000,
            axes_dim=axes_dim_rope,
            axes_lens=axes_lens,
            patch_size=patch_size,
        )

        self.x_embedder = nn.Linear(
            in_features=patch_size * patch_size * in_channels,
            out_features=hidden_size,
        )

        self.ref_image_patch_embedder = nn.Linear(
            in_features=patch_size * patch_size * in_channels,
            out_features=hidden_size,
        )

        self.time_caption_embed = Lumina2CombinedTimestepCaptionEmbedding(
            hidden_size=hidden_size,
            text_feat_dim=text_feat_dim,
            norm_eps=norm_eps,
            timestep_scale=timestep_scale,
        )

        # Initialize transformer blocks
        self.noise_refiner = nn.ModuleList(
            [
                OmniGen2TransformerBlock(
                    hidden_size,
                    num_attention_heads,
                    num_kv_heads,
                    multiple_of,
                    ffn_dim_multiplier,
                    norm_eps,
                    modulation=True,
                    quant_config=quant_config,
                    prefix=f"noise_refiner.{i}",
                )
                for i in range(num_refiner_layers)
            ]
        )

        self.ref_image_refiner = nn.ModuleList(
            [
                OmniGen2TransformerBlock(
                    hidden_size,
                    num_attention_heads,
                    num_kv_heads,
                    multiple_of,
                    ffn_dim_multiplier,
                    norm_eps,
                    modulation=True,
                    quant_config=quant_config,
                    prefix=f"ref_image_refiner.{i}",
                )
                for i in range(num_refiner_layers)
            ]
        )

        self.context_refiner = nn.ModuleList(
            [
                OmniGen2TransformerBlock(
                    hidden_size,
                    num_attention_heads,
                    num_kv_heads,
                    multiple_of,
                    ffn_dim_multiplier,
                    norm_eps,
                    modulation=False,
                    quant_config=quant_config,
                    prefix=f"context_refiner.{i}",
                )
                for i in range(num_refiner_layers)
            ]
        )

        # 3. Transformer blocks
        self.layers = nn.ModuleList(
            [
                OmniGen2TransformerBlock(
                    hidden_size,
                    num_attention_heads,
                    num_kv_heads,
                    multiple_of,
                    ffn_dim_multiplier,
                    norm_eps,
                    modulation=True,
                    quant_config=quant_config,
                    prefix=f"layers.{i}",
                )
                for i in range(num_layers)
            ]
        )

        # 4. Output norm & projection
        self.norm_out = LuminaLayerNormContinuous(
            embedding_dim=hidden_size,
            conditioning_embedding_dim=min(hidden_size, 1024),
            elementwise_affine=False,
            eps=1e-6,
            bias=True,
            out_dim=patch_size * patch_size * self.out_channels,
        )

        # Add learnable embeddings to distinguish different images
        self.image_index_embedding = nn.Parameter(torch.randn(5, hidden_size))  # support max 5 ref images

        self.sp_input_boundary = _OmniGen2SPInputBoundary()
        self.sp_output_boundary = _OmniGen2SPOutputBoundary()

    def img_patch_embed_and_refine(
        self,
        hidden_states,
        ref_image_hidden_states,
        padded_img_mask,
        padded_ref_img_mask,
        noise_rotary_emb,
        ref_img_rotary_emb,
        l_effective_ref_img_len,
        l_effective_img_len,
        temb,
    ):
        batch_size = len(hidden_states)
        has_ref_tokens = any(ref_img_len > 0 for ref_lens in l_effective_ref_img_len for ref_img_len in ref_lens)
        max_combined_img_len = max(
            [img_len + sum(ref_img_len) for img_len, ref_img_len in zip(l_effective_img_len, l_effective_ref_img_len)]
        )

        hidden_states = self.x_embedder(hidden_states)
        if not has_ref_tokens:
            # FP8 kernels do not support zero-token GEMM on ref_image_patch_embedder; skip that path only.
            # Still run noise_refiner and return the same combined layout as the no-ref case below
            # (batch, max_combined_img_len, hidden) — not raw noise tokens alone.
            for layer in self.noise_refiner:
                hidden_states = layer(hidden_states, padded_img_mask, noise_rotary_emb, temb)
            combined_img_hidden_states = hidden_states.new_zeros(
                batch_size, max_combined_img_len, self.config.hidden_size
            )
            for i, img_len in enumerate(l_effective_img_len):
                combined_img_hidden_states[i, :img_len] = hidden_states[i, :img_len]
            return combined_img_hidden_states

        ref_image_hidden_states = self.ref_image_patch_embedder(ref_image_hidden_states)

        for i in range(batch_size):
            shift = 0
            for j, ref_img_len in enumerate(l_effective_ref_img_len[i]):
                ref_image_hidden_states[i, shift : shift + ref_img_len, :] = (
                    ref_image_hidden_states[i, shift : shift + ref_img_len, :] + self.image_index_embedding[j]
                )
                shift += ref_img_len

        for layer in self.noise_refiner:
            hidden_states = layer(hidden_states, padded_img_mask, noise_rotary_emb, temb)

        flat_l_effective_ref_img_len = list(itertools.chain(*l_effective_ref_img_len))
        num_ref_images = len(flat_l_effective_ref_img_len)
        max_ref_img_len = max(flat_l_effective_ref_img_len)

        batch_ref_img_mask = ref_image_hidden_states.new_zeros(num_ref_images, max_ref_img_len, dtype=torch.bool)
        batch_ref_image_hidden_states = ref_image_hidden_states.new_zeros(
            num_ref_images, max_ref_img_len, self.config.hidden_size
        )
        batch_ref_img_rotary_emb = hidden_states.new_zeros(
            num_ref_images,
            max_ref_img_len,
            ref_img_rotary_emb.shape[-1],
            dtype=ref_img_rotary_emb.dtype,
        )
        batch_temb = temb.new_zeros(num_ref_images, *temb.shape[1:], dtype=temb.dtype)

        # sequence of ref imgs to batch
        idx = 0
        for i in range(batch_size):
            shift = 0
            for ref_img_len in l_effective_ref_img_len[i]:
                batch_ref_img_mask[idx, :ref_img_len] = True
                batch_ref_image_hidden_states[idx, :ref_img_len] = ref_image_hidden_states[
                    i, shift : shift + ref_img_len
                ]
                batch_ref_img_rotary_emb[idx, :ref_img_len] = ref_img_rotary_emb[i, shift : shift + ref_img_len]
                batch_temb[idx] = temb[i]
                shift += ref_img_len
                idx += 1

        # refine ref imgs separately
        for layer in self.ref_image_refiner:
            batch_ref_image_hidden_states = layer(
                batch_ref_image_hidden_states,
                batch_ref_img_mask,
                batch_ref_img_rotary_emb,
                batch_temb,
            )

        # batch of ref imgs to sequence
        idx = 0
        for i in range(batch_size):
            shift = 0
            for ref_img_len in l_effective_ref_img_len[i]:
                ref_image_hidden_states[i, shift : shift + ref_img_len] = batch_ref_image_hidden_states[
                    idx, :ref_img_len
                ]
                shift += ref_img_len
                idx += 1

        combined_img_hidden_states = hidden_states.new_zeros(batch_size, max_combined_img_len, self.config.hidden_size)
        for i, (ref_img_len, img_len) in enumerate(zip(l_effective_ref_img_len, l_effective_img_len)):
            combined_img_hidden_states[i, : sum(ref_img_len)] = ref_image_hidden_states[i, : sum(ref_img_len)]
            combined_img_hidden_states[i, sum(ref_img_len) : sum(ref_img_len) + img_len] = hidden_states[i, :img_len]

        return combined_img_hidden_states

    def flat_and_pad_to_seq(self, hidden_states, ref_image_hidden_states):
        batch_size = len(hidden_states)
        p = self.config.patch_size
        device = hidden_states[0].device

        img_sizes = [(img.size(1), img.size(2)) for img in hidden_states]
        l_effective_img_len = [(H // p) * (W // p) for (H, W) in img_sizes]

        if ref_image_hidden_states is not None:
            ref_img_sizes = [
                ([(img.size(1), img.size(2)) for img in imgs] if imgs is not None else None)
                for imgs in ref_image_hidden_states
            ]
            l_effective_ref_img_len = [
                (
                    [(ref_img_size[0] // p) * (ref_img_size[1] // p) for ref_img_size in _ref_img_sizes]
                    if _ref_img_sizes is not None
                    else [0]
                )
                for _ref_img_sizes in ref_img_sizes
            ]
        else:
            ref_img_sizes = [None for _ in range(batch_size)]
            l_effective_ref_img_len = [[0] for _ in range(batch_size)]

        max_ref_img_len = max([sum(ref_img_len) for ref_img_len in l_effective_ref_img_len])
        max_img_len = max(l_effective_img_len)

        # ref image patch embeddings
        flat_ref_img_hidden_states = []
        for i in range(batch_size):
            if ref_img_sizes[i] is not None:
                imgs = []
                for ref_img in ref_image_hidden_states[i]:
                    C, H, W = ref_img.size()
                    ref_img = rearrange(ref_img, "c (h p1) (w p2) -> (h w) (p1 p2 c)", p1=p, p2=p)
                    imgs.append(ref_img)

                img = torch.cat(imgs, dim=0)
                flat_ref_img_hidden_states.append(img)
            else:
                flat_ref_img_hidden_states.append(None)

        # image patch embeddings
        flat_hidden_states = []
        for i in range(batch_size):
            img = hidden_states[i]
            C, H, W = img.size()

            img = rearrange(img, "c (h p1) (w p2) -> (h w) (p1 p2 c)", p1=p, p2=p)
            flat_hidden_states.append(img)

        padded_ref_img_hidden_states = torch.zeros(
            batch_size,
            max_ref_img_len,
            flat_hidden_states[0].shape[-1],
            device=device,
            dtype=flat_hidden_states[0].dtype,
        )
        padded_ref_img_mask = torch.zeros(batch_size, max_ref_img_len, dtype=torch.bool, device=device)
        for i in range(batch_size):
            if ref_img_sizes[i] is not None:
                padded_ref_img_hidden_states[i, : sum(l_effective_ref_img_len[i])] = flat_ref_img_hidden_states[i]
                padded_ref_img_mask[i, : sum(l_effective_ref_img_len[i])] = True

        padded_hidden_states = torch.zeros(
            batch_size,
            max_img_len,
            flat_hidden_states[0].shape[-1],
            device=device,
            dtype=flat_hidden_states[0].dtype,
        )
        padded_img_mask = torch.zeros(batch_size, max_img_len, dtype=torch.bool, device=device)
        for i in range(batch_size):
            padded_hidden_states[i, : l_effective_img_len[i]] = flat_hidden_states[i]
            padded_img_mask[i, : l_effective_img_len[i]] = True

        return (
            padded_hidden_states,
            padded_ref_img_hidden_states,
            padded_img_mask,
            padded_ref_img_mask,
            l_effective_ref_img_len,
            l_effective_img_len,
            ref_img_sizes,
            img_sizes,
        )

    def forward(
        self,
        hidden_states: torch.Tensor | list[torch.Tensor],
        timestep: torch.Tensor,
        text_hidden_states: torch.Tensor,
        freqs_cis: torch.Tensor,
        text_attention_mask: torch.Tensor,
        ref_image_hidden_states: list[list[torch.Tensor]] | None = None,
        return_dict: bool = False,
    ) -> torch.Tensor | Transformer2DModelOutput:
        # 1. Condition, positional & patch embedding
        batch_size = len(hidden_states)
        is_hidden_states_tensor = isinstance(hidden_states, torch.Tensor)

        if is_hidden_states_tensor:
            assert hidden_states.ndim == 4
            hidden_states = [_hidden_states for _hidden_states in hidden_states]

        device = hidden_states[0].device

        temb, text_hidden_states = self.time_caption_embed(timestep, text_hidden_states, hidden_states[0].dtype)

        (
            hidden_states,
            ref_image_hidden_states,
            img_mask,
            ref_img_mask,
            l_effective_ref_img_len,
            l_effective_img_len,
            ref_img_sizes,
            img_sizes,
        ) = self.flat_and_pad_to_seq(hidden_states, ref_image_hidden_states)

        (
            context_rotary_emb,
            ref_img_rotary_emb,
            noise_rotary_emb,
            rotary_emb,
            encoder_seq_lengths,
            seq_lengths,
        ) = self.rope_embedder(
            freqs_cis,
            text_attention_mask,
            l_effective_ref_img_len,
            l_effective_img_len,
            ref_img_sizes,
            img_sizes,
            device,
        )

        # 2. Context refinement
        for layer in self.context_refiner:
            text_hidden_states = layer(text_hidden_states, text_attention_mask, context_rotary_emb)

        combined_img_hidden_states = self.img_patch_embed_and_refine(
            hidden_states,
            ref_image_hidden_states,
            img_mask,
            ref_img_mask,
            noise_rotary_emb,
            ref_img_rotary_emb,
            l_effective_ref_img_len,
            l_effective_img_len,
            temb,
        )

        # 3. Joint Transformer blocks
        max_seq_len = max(seq_lengths)

        attention_mask = hidden_states.new_zeros(batch_size, max_seq_len, dtype=torch.bool)
        joint_hidden_states = hidden_states.new_zeros(batch_size, max_seq_len, self.config.hidden_size)
        for i, (encoder_seq_len, seq_len) in enumerate(zip(encoder_seq_lengths, seq_lengths)):
            attention_mask[i, :seq_len] = True
            joint_hidden_states[i, :encoder_seq_len] = text_hidden_states[i, :encoder_seq_len]
            joint_hidden_states[i, encoder_seq_len:seq_len] = combined_img_hidden_states[i, : seq_len - encoder_seq_len]

        hidden_states = joint_hidden_states

        hidden_states, rotary_emb = self.sp_input_boundary(hidden_states, rotary_emb)

        for layer in self.layers:
            hidden_states = layer(hidden_states, attention_mask, rotary_emb, temb)

        hidden_states = self.sp_output_boundary(hidden_states)

        # 4. Output norm & projection
        hidden_states = self.norm_out(hidden_states, temb)

        p = self.config.patch_size
        output = []
        for i, (img_size, img_len, seq_len) in enumerate(zip(img_sizes, l_effective_img_len, seq_lengths)):
            height, width = img_size
            output.append(
                rearrange(
                    hidden_states[i][seq_len - img_len : seq_len],
                    "(h w) (p1 p2 c) -> c (h p1) (w p2)",
                    h=height // p,
                    w=width // p,
                    p1=p,
                    p2=p,
                )
            )
        if is_hidden_states_tensor:
            output = torch.stack(output, dim=0)

        if not return_dict:
            return output
        return Transformer2DModelOutput(sample=output)

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            # self-attn
            (".to_qkv", ".to_q", "q"),
            (".to_qkv", ".to_k", "k"),
            (".to_qkv", ".to_v", "v"),
            # feed-forward
            (".feed_forward.gate_up_proj", ".feed_forward.linear_1", 0),
            (".feed_forward.gate_up_proj", ".feed_forward.linear_3", 1),
        ]

        params_dict = dict(self.named_parameters())

        loaded_params = set[str]()
        for name, loaded_weight in weights:
            for param_name, weight_name, shard_id in stacked_params_mapping:
                if weight_name not in name:
                    continue
                name = name.replace(weight_name, param_name)
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                # feed_forward.linear_2 -> feed_forward.down_proj rename
                if ".feed_forward.linear_2." in name:
                    name = name.replace(".feed_forward.linear_2.", ".feed_forward.down_proj.")
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader", default_weight_loader)
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params