Meissonic/src/transformer_video.py

import math
from typing import Optional

import torch
import torch.nn as nn
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.models.modeling_utils import ModelMixin

# Global debug flag - set to False to disable debug prints
DEBUG_TRANSFORMER = False

# from .attention import flash_attention
import torch

try:
    import flash_attn_interface
    FLASH_ATTN_3_AVAILABLE = True
except ModuleNotFoundError:
    FLASH_ATTN_3_AVAILABLE = False

try:
    import flash_attn
    FLASH_ATTN_2_AVAILABLE = True
except ModuleNotFoundError:
    FLASH_ATTN_2_AVAILABLE = False

import warnings

__all__ = [
    'flash_attention',
    'attention',
]


def flash_attention(
    q,
    k,
    v,
    q_lens=None,
    k_lens=None,
    dropout_p=0.,
    softmax_scale=None,
    q_scale=None,
    causal=False,
    window_size=(-1, -1),
    deterministic=False,
    dtype=torch.bfloat16,
    version=None,
):
    """
    q:              [B, Lq, Nq, C1].
    k:              [B, Lk, Nk, C1].
    v:              [B, Lk, Nk, C2]. Nq must be divisible by Nk.
    q_lens:         [B].
    k_lens:         [B].
    dropout_p:      float. Dropout probability.
    softmax_scale:  float. The scaling of QK^T before applying softmax.
    causal:         bool. Whether to apply causal attention mask.
    window_size:    (left right). If not (-1, -1), apply sliding window local attention.
    deterministic:  bool. If True, slightly slower and uses more memory.
    dtype:          torch.dtype. Apply when dtype of q/k/v is not float16/bfloat16.
    """
    half_dtypes = (torch.float16, torch.bfloat16)
    assert dtype in half_dtypes
    assert q.device.type == 'cuda' and q.size(-1) <= 256

    # params
    b, lq, lk, out_dtype = q.size(0), q.size(1), k.size(1), q.dtype

    def half(x):
        return x if x.dtype in half_dtypes else x.to(dtype)

    # preprocess query
    if q_lens is None:
        q = half(q.flatten(0, 1))
        q_lens = torch.tensor(
            [lq] * b, dtype=torch.int32).to(
                device=q.device, non_blocking=True)
    else:
        q = half(torch.cat([u[:v] for u, v in zip(q, q_lens)]))

    # preprocess key, value
    if k_lens is None:
        k = half(k.flatten(0, 1))
        v = half(v.flatten(0, 1))
        k_lens = torch.tensor(
            [lk] * b, dtype=torch.int32).to(
                device=k.device, non_blocking=True)
    else:
        k = half(torch.cat([u[:v] for u, v in zip(k, k_lens)]))
        v = half(torch.cat([u[:v] for u, v in zip(v, k_lens)]))

    q = q.to(v.dtype)
    k = k.to(v.dtype)

    if q_scale is not None:
        q = q * q_scale

    if version is not None and version == 3 and not FLASH_ATTN_3_AVAILABLE:
        warnings.warn(
            'Flash attention 3 is not available, use flash attention 2 instead.'
        )

    # apply attention
    if (version is None or version == 3) and FLASH_ATTN_3_AVAILABLE:
        # Note: dropout_p, window_size are not supported in FA3 now.
        x = flash_attn_interface.flash_attn_varlen_func(
            q=q,
            k=k,
            v=v,
            cu_seqlens_q=torch.cat([q_lens.new_zeros([1]), q_lens]).cumsum(
                0, dtype=torch.int32).to(q.device, non_blocking=True),
            cu_seqlens_k=torch.cat([k_lens.new_zeros([1]), k_lens]).cumsum(
                0, dtype=torch.int32).to(q.device, non_blocking=True),
            seqused_q=None,
            seqused_k=None,
            max_seqlen_q=lq,
            max_seqlen_k=lk,
            softmax_scale=softmax_scale,
            causal=causal,
            deterministic=deterministic)[0].unflatten(0, (b, lq))
    else:
        assert FLASH_ATTN_2_AVAILABLE
        x = flash_attn.flash_attn_varlen_func(
            q=q,
            k=k,
            v=v,
            cu_seqlens_q=torch.cat([q_lens.new_zeros([1]), q_lens]).cumsum(
                0, dtype=torch.int32).to(q.device, non_blocking=True),
            cu_seqlens_k=torch.cat([k_lens.new_zeros([1]), k_lens]).cumsum(
                0, dtype=torch.int32).to(q.device, non_blocking=True),
            max_seqlen_q=lq,
            max_seqlen_k=lk,
            dropout_p=dropout_p,
            softmax_scale=softmax_scale,
            causal=causal,
            window_size=window_size,
            deterministic=deterministic).unflatten(0, (b, lq))

    # output
    return x.type(out_dtype)


def attention(
    q,
    k,
    v,
    q_lens=None,
    k_lens=None,
    dropout_p=0.,
    softmax_scale=None,
    q_scale=None,
    causal=False,
    window_size=(-1, -1),
    deterministic=False,
    dtype=torch.bfloat16,
    fa_version=None,
):
    if FLASH_ATTN_2_AVAILABLE or FLASH_ATTN_3_AVAILABLE:
        return flash_attention(
            q=q,
            k=k,
            v=v,
            q_lens=q_lens,
            k_lens=k_lens,
            dropout_p=dropout_p,
            softmax_scale=softmax_scale,
            q_scale=q_scale,
            causal=causal,
            window_size=window_size,
            deterministic=deterministic,
            dtype=dtype,
            version=fa_version,
        )
    else:
        if q_lens is not None or k_lens is not None:
            warnings.warn(
                'Padding mask is disabled when using scaled_dot_product_attention. It can have a significant impact on performance.'
            )
        attn_mask = None

        q = q.transpose(1, 2).to(dtype)
        k = k.transpose(1, 2).to(dtype)
        v = v.transpose(1, 2).to(dtype)

        out = torch.nn.functional.scaled_dot_product_attention(
            q, k, v, attn_mask=attn_mask, is_causal=causal, dropout_p=dropout_p)

        out = out.transpose(1, 2).contiguous()
        return out


__all__ = ['WanModel']


def sinusoidal_embedding_1d(dim, position):
    # preprocess
    assert dim % 2 == 0
    half = dim // 2
    # Ensure position is on CPU for float64 computation to avoid CUDA issues
    # Convert to float64 for precision, then move back to original device
    device = position.device
    position = position.to(torch.float64)

    # calculation
    # Create range tensor on same device as position
    arange_tensor = torch.arange(half, dtype=torch.float64, device=device)
    sinusoid = torch.outer(
        position, torch.pow(10000, -arange_tensor.div(half)))
    x = torch.cat([torch.cos(sinusoid), torch.sin(sinusoid)], dim=1)
    return x


@torch.amp.autocast('cuda', enabled=False)
def rope_params(max_seq_len, dim, theta=10000):
    assert dim % 2 == 0
    freqs = torch.outer(
        torch.arange(max_seq_len),
        1.0 / torch.pow(theta,
                        torch.arange(0, dim, 2).to(torch.float64).div(dim)))
    freqs = torch.polar(torch.ones_like(freqs), freqs)
    return freqs


@torch.amp.autocast('cuda', enabled=False)
def rope_apply(x, grid_sizes, freqs):
    n, c = x.size(2), x.size(3) // 2
    # Save original dtype to restore it later
    original_dtype = x.dtype

    # split freqs
    freqs = freqs.split([c - 2 * (c // 3), c // 3, c // 3], dim=1)

    # loop over samples
    output = []
    for i, (f, h, w) in enumerate(grid_sizes.tolist()):
        seq_len = f * h * w

        # precompute multipliers
        x_i = torch.view_as_complex(x[i, :seq_len].to(torch.float64).reshape(
            seq_len, n, -1, 2))
        freqs_i = torch.cat([
            freqs[0][:f].view(f, 1, 1, -1).expand(f, h, w, -1),
            freqs[1][:h].view(1, h, 1, -1).expand(f, h, w, -1),
            freqs[2][:w].view(1, 1, w, -1).expand(f, h, w, -1)
        ],
                            dim=-1).reshape(seq_len, 1, -1)

        # apply rotary embedding
        x_i = torch.view_as_real(x_i * freqs_i).flatten(2)
        # Convert back to original dtype before concatenating
        x_i = x_i.to(dtype=original_dtype)
        # Handle the remaining part of the sequence
        x_remaining = x[i, seq_len:]
        if x_remaining.numel() > 0:
            x_i = torch.cat([x_i, x_remaining])
        else:
            x_i = x_i

        # append to collection
        output.append(x_i)
    # Stack and ensure dtype matches original input
    return torch.stack(output).to(dtype=original_dtype)


class WanRMSNorm(nn.Module):

    def __init__(self, dim, eps=1e-5):
        super().__init__()
        self.dim = dim
        self.eps = eps
        self.weight = nn.Parameter(torch.ones(dim))

    def forward(self, x):
        r"""
        Args:
            x(Tensor): Shape [B, L, C]
        """
        # Ensure weight dtype matches input dtype
        return self._norm(x.float()).type_as(x) * self.weight.type_as(x)

    def _norm(self, x):
        return x * torch.rsqrt(x.pow(2).mean(dim=-1, keepdim=True) + self.eps)


class WanLayerNorm(nn.LayerNorm):

    def __init__(self, dim, eps=1e-6, elementwise_affine=False):
        super().__init__(dim, elementwise_affine=elementwise_affine, eps=eps)

    def forward(self, x):
        r"""
        Args:
            x(Tensor): Shape [B, L, C]
        """
        # Convert to float32 for numerical stability, ensuring weights match input dtype
        original_dtype = x.dtype
        x_float = x.float()
        if self.elementwise_affine:
            weight_float = self.weight.float() if self.weight is not None else None
            bias_float = self.bias.float() if self.bias is not None else None
            # Use torch.nn.functional.layer_norm directly with converted weights
            result = torch.nn.functional.layer_norm(x_float, self.normalized_shape, weight_float, bias_float, self.eps)
        else:
            result = super().forward(x_float)
        return result.to(dtype=original_dtype)


class WanSelfAttention(nn.Module):

    def __init__(self,
                 dim,
                 num_heads,
                 window_size=(-1, -1),
                 qk_norm=True,
                 eps=1e-6):
        assert dim % num_heads == 0
        super().__init__()
        self.dim = dim
        self.num_heads = num_heads
        self.head_dim = dim // num_heads
        self.window_size = window_size
        self.qk_norm = qk_norm
        self.eps = eps

        # layers
        self.q = nn.Linear(dim, dim)
        self.k = nn.Linear(dim, dim)
        self.v = nn.Linear(dim, dim)
        self.o = nn.Linear(dim, dim)
        self.norm_q = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity()
        self.norm_k = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity()

    def forward(self, x, seq_lens, grid_sizes, freqs):
        r"""
        Args:
            x(Tensor): Shape [B, L, num_heads, C / num_heads]
            seq_lens(Tensor): Shape [B]
            grid_sizes(Tensor): Shape [B, 3], the second dimension contains (F, H, W)
            freqs(Tensor): Rope freqs, shape [1024, C / num_heads / 2]
        """
        b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim

        # query, key, value function
        def qkv_fn(x):
            q = self.norm_q(self.q(x)).view(b, s, n, d)
            k = self.norm_k(self.k(x)).view(b, s, n, d)
            v = self.v(x).view(b, s, n, d)
            return q, k, v

        q, k, v = qkv_fn(x)
        
        # Save input dtype to ensure output matches
        input_dtype = x.dtype

        x = flash_attention(
            q=rope_apply(q, grid_sizes, freqs),
            k=rope_apply(k, grid_sizes, freqs),
            v=v,
            k_lens=seq_lens,
            window_size=self.window_size)
        
        # Ensure output dtype matches input dtype (in case rope_apply or flash_attention changed it)
        x = x.to(dtype=input_dtype)

        # output
        x = x.flatten(2)
        x = self.o(x)
        return x


class WanCrossAttention(WanSelfAttention):

    def forward(self, x, context, context_lens):
        r"""
        Args:
            x(Tensor): Shape [B, L1, C]
            context(Tensor): Shape [B, L2, C]
            context_lens(Tensor): Shape [B]
        """
        b, n, d = x.size(0), self.num_heads, self.head_dim
        
        # Save input dtype to ensure output matches
        input_dtype = x.dtype

        # compute query, key, value
        q = self.norm_q(self.q(x)).view(b, -1, n, d)
        k = self.norm_k(self.k(context)).view(b, -1, n, d)
        v = self.v(context).view(b, -1, n, d)

        # compute attention
        x = flash_attention(q, k, v, k_lens=context_lens)
        
        # Ensure output dtype matches input dtype
        x = x.to(dtype=input_dtype)

        # output
        x = x.flatten(2)
        x = self.o(x)
        return x


class WanAttentionBlock(nn.Module):

    def __init__(self,
                 dim,
                 ffn_dim,
                 num_heads,
                 window_size=(-1, -1),
                 qk_norm=True,
                 cross_attn_norm=False,
                 eps=1e-6):
        super().__init__()
        self.dim = dim
        self.ffn_dim = ffn_dim
        self.num_heads = num_heads
        self.window_size = window_size
        self.qk_norm = qk_norm
        self.cross_attn_norm = cross_attn_norm
        self.eps = eps

        # layers
        self.norm1 = WanLayerNorm(dim, eps)
        self.self_attn = WanSelfAttention(dim, num_heads, window_size, qk_norm,
                                          eps)
        self.norm3 = WanLayerNorm(
            dim, eps,
            elementwise_affine=True) if cross_attn_norm else nn.Identity()
        self.cross_attn = WanCrossAttention(dim, num_heads, (-1, -1), qk_norm,
                                            eps)
        self.norm2 = WanLayerNorm(dim, eps)
        self.ffn = nn.Sequential(
            nn.Linear(dim, ffn_dim), nn.GELU(approximate='tanh'),
            nn.Linear(ffn_dim, dim))

        # modulation
        self.modulation = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5)

    def forward(
        self,
        x,
        e,
        seq_lens,
        grid_sizes,
        freqs,
        context,
        context_lens,
    ):
        r"""
        Args:
            x(Tensor): Shape [B, L, C]
            e(Tensor): Shape [B, L1, 6, C]
            seq_lens(Tensor): Shape [B], length of each sequence in batch
            grid_sizes(Tensor): Shape [B, 3], the second dimension contains (F, H, W)
            freqs(Tensor): Rope freqs, shape [1024, C / num_heads / 2]
        """
        # Convert e to float32 for modulation computation (modulation expects float32)
        e_float32 = e.to(dtype=torch.float32) if e.dtype != torch.float32 else e
        with torch.amp.autocast('cuda', dtype=torch.float32):
            e = (self.modulation.unsqueeze(0) + e_float32).chunk(6, dim=2)
        assert e[0].dtype == torch.float32

        # self-attention
        # Ensure input dtype matches model weights (convert e to match x's dtype)
        x_dtype = x.dtype
        e_0 = e[0].squeeze(2).to(dtype=x_dtype)
        e_1 = e[1].squeeze(2).to(dtype=x_dtype)
        e_2 = e[2].squeeze(2).to(dtype=x_dtype)
        attn_input = self.norm1(x) * (1 + e_1) + e_0
        y = self.self_attn(attn_input, seq_lens, grid_sizes, freqs)
        # Ensure dtype consistency: y and e_2 should match x's dtype
        x = x + (y * e_2).to(dtype=x_dtype)

        # cross-attention & ffn function
        def cross_attn_ffn(x, context, context_lens, e):
            x = x + self.cross_attn(self.norm3(x), context, context_lens)
            # Ensure dtype consistency for FFN input
            x_dtype = x.dtype
            e_3 = e[3].squeeze(2).to(dtype=x_dtype)
            e_4 = e[4].squeeze(2).to(dtype=x_dtype)
            e_5 = e[5].squeeze(2).to(dtype=x_dtype)
            ffn_input = self.norm2(x) * (1 + e_4) + e_3
            y = self.ffn(ffn_input)
            # Ensure dtype consistency: y and e_5 should match x's dtype
            x = x + (y * e_5).to(dtype=x_dtype)
            return x

        x = cross_attn_ffn(x, context, context_lens, e)
        return x


class Head(nn.Module):

    def __init__(self, dim, out_dim, patch_size, eps=1e-6):
        super().__init__()
        self.dim = dim
        self.out_dim = out_dim
        self.patch_size = patch_size
        self.eps = eps

        # layers
        out_dim = math.prod(patch_size) * out_dim
        self.norm = WanLayerNorm(dim, eps)
        self.head = nn.Linear(dim, out_dim)

        # modulation
        self.modulation = nn.Parameter(torch.randn(1, 2, dim) / dim**0.5)

    def forward(self, x, e):
        r"""
        Args:
            x(Tensor): Shape [B, L1, C]
            e(Tensor): Shape [B, L1, C]
        """
        # Convert e to float32 for modulation computation (modulation expects float32)
        e_float32 = e.to(dtype=torch.float32) if e.dtype != torch.float32 else e
        with torch.amp.autocast('cuda', dtype=torch.float32):
            e = (self.modulation.unsqueeze(0) + e_float32.unsqueeze(2)).chunk(2, dim=2)
        # Ensure dtype consistency: convert e to match x's dtype
        x_dtype = x.dtype
        e_0 = e[0].squeeze(2).to(dtype=x_dtype)
        e_1 = e[1].squeeze(2).to(dtype=x_dtype)
        head_input = self.norm(x) * (1 + e_1) + e_0
        x = self.head(head_input)
        return x


class WanModel(ModelMixin, ConfigMixin):
    r"""
    Wan diffusion backbone supporting both text-to-video and image-to-video.
    """

    ignore_for_config = [
        'patch_size', 'cross_attn_norm', 'qk_norm', 'text_dim', 'window_size'
    ]
    _no_split_modules = ['WanAttentionBlock']

    @register_to_config
    def __init__(self,
                 model_type='t2v',
                 patch_size=(1, 2, 2),
                 text_len=512,
                 in_dim=16,
                 dim=2048,
                 ffn_dim=8192,
                 freq_dim=256,
                 text_dim=4096,
                 out_dim=16,
                 num_heads=16,
                 num_layers=32,
                 window_size=(-1, -1),
                 qk_norm=True,
                 cross_attn_norm=True,
                 eps=1e-6):
        r"""
        Initialize the diffusion model backbone.

        Args:
            model_type (`str`, *optional*, defaults to 't2v'):
                Model variant - 't2v' (text-to-video) or 'i2v' (image-to-video)
            patch_size (`tuple`, *optional*, defaults to (1, 2, 2)):
                3D patch dimensions for video embedding (t_patch, h_patch, w_patch)
            text_len (`int`, *optional*, defaults to 512):
                Fixed length for text embeddings
            in_dim (`int`, *optional*, defaults to 16):
                Input video channels (C_in)
            dim (`int`, *optional*, defaults to 2048):
                Hidden dimension of the transformer
            ffn_dim (`int`, *optional*, defaults to 8192):
                Intermediate dimension in feed-forward network
            freq_dim (`int`, *optional*, defaults to 256):
                Dimension for sinusoidal time embeddings
            text_dim (`int`, *optional*, defaults to 4096):
                Input dimension for text embeddings
            out_dim (`int`, *optional*, defaults to 16):
                Output video channels (C_out)
            num_heads (`int`, *optional*, defaults to 16):
                Number of attention heads
            num_layers (`int`, *optional*, defaults to 32):
                Number of transformer blocks
            window_size (`tuple`, *optional*, defaults to (-1, -1)):
                Window size for local attention (-1 indicates global attention)
            qk_norm (`bool`, *optional*, defaults to True):
                Enable query/key normalization
            cross_attn_norm (`bool`, *optional*, defaults to False):
                Enable cross-attention normalization
            eps (`float`, *optional*, defaults to 1e-6):
                Epsilon value for normalization layers
        """

        super().__init__()

        assert model_type in ['t2v', 'i2v', 'ti2v', 's2v']
        self.model_type = model_type

        self.patch_size = patch_size
        self.text_len = text_len
        self.in_dim = in_dim
        self.dim = dim
        self.ffn_dim = ffn_dim
        self.freq_dim = freq_dim
        self.text_dim = text_dim
        self.out_dim = out_dim
        self.num_heads = num_heads
        self.num_layers = num_layers
        self.window_size = window_size
        self.qk_norm = qk_norm
        self.cross_attn_norm = cross_attn_norm
        self.eps = eps

        # embeddings
        self.patch_embedding = nn.Conv3d(
            in_dim, dim, kernel_size=patch_size, stride=patch_size)
        self.text_embedding = nn.Sequential(
            nn.Linear(text_dim, dim), nn.GELU(approximate='tanh'),
            nn.Linear(dim, dim))

        self.time_embedding = nn.Sequential(
            nn.Linear(freq_dim, dim), nn.SiLU(), nn.Linear(dim, dim))
        self.time_projection = nn.Sequential(nn.SiLU(), nn.Linear(dim, dim * 6))

        # blocks
        self.blocks = nn.ModuleList([
            WanAttentionBlock(dim, ffn_dim, num_heads, window_size, qk_norm,
                              cross_attn_norm, eps) for _ in range(num_layers)
        ])

        # head
        self.head = Head(dim, out_dim, patch_size, eps)

        # buffers (don't use register_buffer otherwise dtype will be changed in to())
        assert (dim % num_heads) == 0 and (dim // num_heads) % 2 == 0
        d = dim // num_heads
        self.freqs = torch.cat([
            rope_params(1024, d - 4 * (d // 6)),
            rope_params(1024, 2 * (d // 6)),
            rope_params(1024, 2 * (d // 6))
        ],
                               dim=1)

        # initialize weights
        self.init_weights()

    def forward(
        self,
        x,
        t,
        context,
        seq_len,
        y=None,
    ):
        r"""
        Forward pass through the diffusion model

        Args:
            x (List[Tensor]):
                List of input video tensors, each with shape [C_in, F, H, W]
            t (Tensor):
                Diffusion timesteps tensor of shape [B]
            context (List[Tensor]):
                List of text embeddings each with shape [L, C]
            seq_len (`int`):
                Maximum sequence length for positional encoding
            y (List[Tensor], *optional*):
                Conditional video inputs for image-to-video mode, same shape as x

        Returns:
            List[Tensor]:
                List of denoised video tensors with original input shapes [C_out, F, H / 8, W / 8]
        """
        if self.model_type == 'i2v':
            assert y is not None
        # params
        device = self.patch_embedding.weight.device
        if self.freqs.device != device:
            self.freqs = self.freqs.to(device)

        if y is not None:
            x = [torch.cat([u, v], dim=0) for u, v in zip(x, y)]

        # embeddings
        # Ensure input dtype matches patch_embedding weight dtype
        patch_weight_dtype = self.patch_embedding.weight.dtype
        x = [self.patch_embedding(u.unsqueeze(0).to(dtype=patch_weight_dtype)) for u in x]
        grid_sizes = torch.stack(
            [torch.tensor(u.shape[2:], dtype=torch.long) for u in x])
        x = [u.flatten(2).transpose(1, 2) for u in x]
        seq_lens = torch.tensor([u.size(1) for u in x], dtype=torch.long)
        assert seq_lens.max() <= seq_len
        x = torch.cat([
            torch.cat([u, u.new_zeros(1, seq_len - u.size(1), u.size(2))],
                      dim=1) for u in x
        ])
        
        # time embeddings
        if t.dim() == 1:
            t = t.expand(t.size(0), seq_len)
        with torch.amp.autocast('cuda', dtype=torch.float32):
            bt = t.size(0)
            t = t.flatten()
            e = self.time_embedding(
                sinusoidal_embedding_1d(self.freq_dim,
                                        t).unflatten(0, (bt, seq_len)).float())
            e0 = self.time_projection(e).unflatten(2, (6, self.dim))
            assert e.dtype == torch.float32 and e0.dtype == torch.float32
        
        # Keep e and e0 as float32 for modulation computation
        # They will be converted to x.dtype inside WanAttentionBlock.forward and Head.forward when needed

        # context
        context_lens = None
        # Ensure context input dtype matches text_embedding weight dtype
        text_weight_dtype = self.text_embedding[0].weight.dtype
        context = self.text_embedding(
            torch.stack([
                torch.cat(
                    [u, u.new_zeros(self.text_len - u.size(0), u.size(1))])
                for u in context
            ]).to(dtype=text_weight_dtype))

        # arguments
        kwargs = dict(
            e=e0,
            seq_lens=seq_lens,
            grid_sizes=grid_sizes,
            freqs=self.freqs,
            context=context,
            context_lens=context_lens)

        for block in self.blocks:
            x = block(x, **kwargs)

        # head
        x = self.head(x, e)

        # unpatchify
        x = self.unpatchify(x, grid_sizes)
        return [u.float() for u in x]

    def unpatchify(self, x, grid_sizes):
        r"""
        Reconstruct video tensors from patch embeddings.

        Args:
            x (List[Tensor]):
                List of patchified features, each with shape [L, C_out * prod(patch_size)]
            grid_sizes (Tensor):
                Original spatial-temporal grid dimensions before patching,
                    shape [B, 3] (3 dimensions correspond to F_patches, H_patches, W_patches)

        Returns:
            List[Tensor]:
                Reconstructed video tensors with shape [C_out, F, H / 8, W / 8]
        """

        c = self.out_dim
        out = []
        for u, v in zip(x, grid_sizes.tolist()):
            u = u[:math.prod(v)].view(*v, *self.patch_size, c)
            u = torch.einsum('fhwpqrc->cfphqwr', u)
            u = u.reshape(c, *[i * j for i, j in zip(v, self.patch_size)])
            out.append(u)
        return out

    def init_weights(self):
        r"""
        Initialize model parameters using Xavier initialization.
        """

        # basic init
        for m in self.modules():
            if isinstance(m, nn.Linear):
                nn.init.xavier_uniform_(m.weight)
                if m.bias is not None:
                    nn.init.zeros_(m.bias)

        # init embeddings
        nn.init.xavier_uniform_(self.patch_embedding.weight.flatten(1))
        for m in self.text_embedding.modules():
            if isinstance(m, nn.Linear):
                nn.init.normal_(m.weight, std=.02)
        for m in self.time_embedding.modules():
            if isinstance(m, nn.Linear):
                nn.init.normal_(m.weight, std=.02)

        # init output layer
        nn.init.zeros_(self.head.head.weight)




class WanDiscreteVideoTransformer(ModelMixin, ConfigMixin):
    r"""
    Wrapper around :class:`WanModel` that makes it usable as a **discrete video diffusion backbone**.

    The goals of this wrapper are:

    - keep the inner :class:`WanModel` architecture and parameter names intact so that Wan-1.3B
      weights can later be loaded directly into ``self.backbone``;
    - expose a simpler interface that takes **discrete codebook indices** (from a 2D VQ-VAE on
      pseudo-video) and returns **logits over the codebook** for each spatio‑temporal position.

    Notes
    -----
    - This class does **not** try to be drop‑in compatible with Meissonic's 2D ``Transformer2DModel``.
      It is a parallel, video‑oriented path that still follows the same *discrete diffusion* principle:
      predict per‑token logits given masked tokens + text.
    - Pseudo‑video is represented as a 4D integer tensor ``[B, F, H, W]`` of codebook indices.
      How to get these tokens from the current 2D VQ-VAE (e.g. per‑frame encoding & stacking)
      is left to the higher‑level training / pipeline code.
    """

    _supports_gradient_checkpointing = True

    @register_to_config
    def __init__(
        self,
        # discrete codebook settings
        codebook_size: int,
        vocab_size: int,
        # video layout
        num_frames: int,
        height: int,
        width: int,
        # Wan backbone hyper‑parameters (mirrors WanModel.__init__)
        model_type: str = 't2v',
        patch_size: tuple = (1, 2, 2),
        text_len: int = 512,
        in_dim: int = 16,
        dim: int = 2048,
        ffn_dim: int = 8192,
        freq_dim: int = 256,
        text_dim: int = 4096,
        out_dim: int = 16,
        num_heads: int = 16,
        num_layers: int = 32,
        window_size: tuple = (-1, -1),
        qk_norm: bool = True,
        cross_attn_norm: bool = True,
        eps: float = 1e-6,
    ):
        super().__init__()

        # save a minimal set of attributes useful for downstream tooling
        self.codebook_size = codebook_size
        self.vocab_size = vocab_size
        self.num_frames = num_frames
        self.height = height
        self.width = width

        # 1) backbone: keep WanModel intact for future weight loading
        self.backbone = WanModel(
            model_type=model_type,
            patch_size=patch_size,
            text_len=text_len,
            in_dim=in_dim,
            dim=dim,
            ffn_dim=ffn_dim,
            freq_dim=freq_dim,
            text_dim=text_dim,
            out_dim=out_dim,
            num_heads=num_heads,
            num_layers=num_layers,
            window_size=window_size,
            qk_norm=qk_norm,
            cross_attn_norm=cross_attn_norm,
            eps=eps,
        )

        # 2) discrete token embedding -> continuous video volume
        #
        #    Input:  tokens [B, F, H, W] with values in [0, vocab_size) where:
        #            - [0, codebook_size-1] = actual Cosmos codes (direct mapping, no shift)
        #            - codebook_size = mask_token_id (reserved for masking)
        #    Output: list of length B with tensors [in_dim, F, H, W]
        #
        #    We keep this outside the backbone so that loading official Wan 1.3B weights
        #    into self.backbone will still work without clashes.
        #    Note: vocab_size = codebook_size + 1 to accommodate mask_token_id = codebook_size
        self.token_embedding = nn.Embedding(vocab_size, in_dim)

        # 3) projection from continuous video output -> logits over codebook
        #
        #    Backbone output: list of B tensors [out_dim, F, H', W']
        #    We map it with a 3D 1x1x1 conv to [vocab_size, F, H', W'].
        #    Note: vocab_size = codebook_size + 1, where codebook_size is reserved for mask_token_id
        self.logits_head = nn.Conv3d(out_dim, vocab_size, kernel_size=1)
        
        # Gradient checkpointing support
        self.gradient_checkpointing = False

    def _tokens_to_video(self, tokens: torch.LongTensor) -> list:
        r"""
        Convert discrete tokens ``[B, F, H, W]`` into a list of length ``B`` where each element
        is a dense video tensor ``[in_dim, F, H, W]`` suitable for :class:`WanModel`.
        
        Note:
            This method now supports dynamic input dimensions. The num_frames, height, width
            stored in config are used as defaults/for seq_len calculation, but inputs can
            have different dimensions as long as they're valid.
        """
        assert tokens.dim() == 4, f"expected [B, F, H, W] tokens, got {tokens.shape}"
        # Dynamic dimensions - no strict dimension checks, WanModel handles variable sizes

        # [B, F, H, W, in_dim]
        # Ensure output dtype matches token_embedding weight dtype
        x = self.token_embedding(tokens)
        # Ensure dtype matches model's expected dtype (usually bfloat16 for mixed precision)
        token_embedding_dtype = self.token_embedding.weight.dtype
        x = x.to(dtype=token_embedding_dtype)
        # [B, in_dim, F, H, W]
        x = x.permute(0, 4, 1, 2, 3).contiguous()

        # WanModel expects a list of [C_in, F, H, W]
        return [x_i for x_i in x]

    def _text_to_list(self, encoder_hidden_states: torch.Tensor) -> list:
        r"""
        Convert batched text embeddings ``[B, L, C]`` into the list-of-tensors format
        expected by :class:`WanModel`.
        """
        assert encoder_hidden_states.dim() == 3, (
            f"expected encoder_hidden_states [B, L, C], got {encoder_hidden_states.shape}")
        return [e for e in encoder_hidden_states]

    def _set_gradient_checkpointing(self, enable=True, gradient_checkpointing_func=None):
        """Set gradient checkpointing for the module."""
        self.gradient_checkpointing = enable

    def forward(
        self,
        tokens: torch.LongTensor,
        timesteps: torch.LongTensor,
        encoder_hidden_states: torch.FloatTensor,
        y: Optional[list] = None,
    ) -> torch.FloatTensor:
        r"""
        Forward pass of the **discrete video transformer**.

        Args:
            tokens (`torch.LongTensor` of shape `[B, F, H, W]`):
                Discrete codebook indices (e.g. from a 2D VQ-VAE applied frame‑wise).
            timesteps (`torch.LongTensor` of shape `[B]` or `[B, F * H * W]`):
                Diffusion timestep(s), following the same semantics as Meissonic's scalar timesteps.
            encoder_hidden_states (`torch.FloatTensor` of shape `[B, L, C_text]`):
                Text embeddings (e.g. from CLIP). Each sample corresponds to one video.
            y (`Optional[list]`):
                Optional conditional video list passed to the underlying :class:`WanModel`
                for i2v / ti2v / s2v variants. For now this is surfaced as a raw passthrough
                and can be left as ``None`` for pure text‑to‑video.

        Returns:
            `torch.FloatTensor`:
                Logits over the codebook of shape `[B, codebook_size, F, H_out, W_out]`, where
                `(H_out, W_out)` depend on the Wan patch configuration. For the default
                `patch_size=(1, 2, 2)` and input ``H=W=height``, we have
                ``H_out = height // 2`` and ``W_out = width // 2``.
        """
        device = tokens.device
        if DEBUG_TRANSFORMER:
            print(f"[DEBUG-transformer] Input: tokens.shape={tokens.shape}, encoder_hidden_states.shape={encoder_hidden_states.shape}, timesteps.shape={timesteps.shape}")
        x_list = self._tokens_to_video(tokens)
        context_list = self._text_to_list(encoder_hidden_states)
        if DEBUG_TRANSFORMER:
            print(f"[DEBUG-transformer] After conversion: len(x_list)={len(x_list)}, len(context_list)={len(context_list)}")
            if len(x_list) > 0:
                print(f"[DEBUG-transformer] x_list[0].shape={x_list[0].shape}")
            if len(context_list) > 0:
                print(f"[DEBUG-transformer] context_list[0].shape={context_list[0].shape}")

        # Calculate seq_len from actual input dimensions (supports dynamic sizes)
        # tokens: [B, F, H, W] -> after patchification: seq_len = F * (H/p_h) * (W/p_w)
        _, f_in, h_in, w_in = tokens.shape
        h_patch = h_in // self.backbone.patch_size[1]
        w_patch = w_in // self.backbone.patch_size[2]
        seq_len = f_in * h_patch * w_patch

        # Prepare timesteps in the exact shape WanModel.forward expects.
        # Its current implementation assumes `t` is either [B, seq_len] or will be
        # expanded from 1D; the 1D branch is slightly buggy for non-singleton dims,
        # so we always give it a [B, seq_len] tensor here.
        if timesteps.dim() == 1:
            # [B] -> [B, 1] -> [B, seq_len] (broadcast along sequence)
            t_model = timesteps.to(device).unsqueeze(1).expand(-1, seq_len)
        elif timesteps.dim() == 2:
            assert timesteps.size(1) == seq_len, (
                f"Expected timesteps second dim == seq_len ({seq_len}), "
                f"but got {timesteps.size(1)}"
            )
            t_model = timesteps.to(device)
        else:
            raise ValueError(
                f"Unsupported timesteps shape {timesteps.shape}; "
                "expected [B] or [B, seq_len]"
            )
        if DEBUG_TRANSFORMER:
            print(f"[DEBUG-transformer] t_model.shape={t_model.shape}")

        # WanModel.forward expects:
        #   x: List[Tensor [C_in, F, H, W]]
        #   t: Tensor [B] or [B, seq_len]
        #   context: List[Tensor [L, C_text]]
        #   seq_len: int
        #   y: Optional[List[Tensor]]
        if self.training and self.gradient_checkpointing:
            def create_custom_forward(module):
                def custom_forward(*inputs):
                    # Unpack inputs: x_list, t, context_list, seq_len, y
                    x_in, t_in, context_in, seq_len_in, y_in = inputs
                    return module(x=x_in, t=t_in, context=context_in, seq_len=seq_len_in, y=y_in)
                return custom_forward
            
            # Use gradient checkpointing for the backbone
            ckpt_kwargs = {"use_reentrant": False}
            out_list = torch.utils.checkpoint.checkpoint(
                create_custom_forward(self.backbone),
                x_list,
                t_model,
                context_list,
                seq_len,
                y,
                **ckpt_kwargs,
            )
        else:
            out_list = self.backbone(
                x=x_list,
                t=t_model,
                context=context_list,
                seq_len=seq_len,
                y=y,
            )
        if DEBUG_TRANSFORMER:
            print(f"[DEBUG-transformer] After backbone: len(out_list)={len(out_list)}")
            if len(out_list) > 0:
                print(f"[DEBUG-transformer] out_list[0].shape={out_list[0].shape}")

        # out_list: length B, each [C_out, F, H_out, W_out]
        vids = torch.stack(out_list, dim=0)  # [B, C_out, F, H_out, W_out]
        if DEBUG_TRANSFORMER:
            print(f"[DEBUG-transformer] After stack: vids.shape={vids.shape}")
        # Ensure vids dtype matches logits_head weight dtype
        vids = vids.to(dtype=self.logits_head.weight.dtype)
        logits = self.logits_head(vids)  # [B, vocab_size, F, H_out, W_out] where vocab_size = codebook_size + 1
        if DEBUG_TRANSFORMER:
            print(f"[DEBUG-transformer] Final logits.shape={logits.shape}")
        return logits
    
# def _available_device():
#     return "cuda" if torch.cuda.is_available() else "cpu"


# def test_wan_discrete_video_transformer_forward_and_shapes():
#     """
#     Basic smoke test:
#     - build a tiny WanDiscreteVideoTransformer
#     - run a forward pass with random pseudo-video tokens + random text
#     - check output shapes, parameter count and (if CUDA present) memory usage
#     """

#     device = _available_device()

#     # small config to keep the test lightweight
#     codebook_size = 128
#     vocab_size = codebook_size + 1  # reserve one for mask if needed later
#     num_frames = 2
#     height = 16
#     width = 16

#     model = WanDiscreteVideoTransformer(
#         codebook_size=codebook_size,
#         vocab_size=vocab_size,
#         num_frames=num_frames,
#         height=height,
#         width=width,
#         # shrink Wan backbone for the unit test
#         in_dim=32,
#         dim=64,
#         ffn_dim=128,
#         freq_dim=32,
#         text_dim=64,
#         out_dim=32,
#         num_heads=4,
#         num_layers=2,
#     ).to(device)
#     model.eval()

#     batch_size = 2

#     # pseudo-video tokens from 2D VQ-VAE on frames: [B, F, H, W]
#     tokens = torch.randint(
#         low=0,
#         high=codebook_size,
#         size=(batch_size, num_frames, height, width),
#         dtype=torch.long,
#         device=device,
#     )

#     # text: [B, L, C_text]
#     text_seq_len = 8
#     encoder_hidden_states = torch.randn(
#         batch_size, text_seq_len, model.backbone.text_dim, device=device
#     )

#     # timesteps: [B]
#     timesteps = torch.randint(
#         low=0, high=1000, size=(batch_size,), dtype=torch.long, device=device
#     )

#     # track memory if CUDA is available
#     if device == "cuda":
#         torch.cuda.reset_peak_memory_stats()
#         mem_before = torch.cuda.memory_allocated()
#     else:
#         mem_before = 0

#     with torch.no_grad():
#         logits = model(
#             tokens=tokens,
#             timesteps=timesteps,
#             encoder_hidden_states=encoder_hidden_states,
#             y=None,
#         )

#     if device == "cuda":
#         mem_after = torch.cuda.memory_allocated()
#         peak_mem = torch.cuda.max_memory_allocated()
#     else:
#         mem_after = mem_before
#         peak_mem = mem_before

#     # logits: [B, codebook_size, F, H_out, W_out]
#     assert logits.shape[0] == batch_size
#     assert logits.shape[1] == codebook_size
#     assert logits.shape[2] == num_frames

#      # WanModel returns unpatchified videos, so spatial size matches the input grid.
#     h_out = height
#     w_out = width
#     assert logits.shape[3] == h_out
#     assert logits.shape[4] == w_out

#     # parameter count sanity check (just ensure it's > 0 and finite)
#     num_params = sum(p.numel() for p in model.parameters())
#     assert num_params > 0
#     assert math.isfinite(float(num_params))

#     # memory sanity check (on CUDA the forward pass should allocate > 0 bytes)
#     if device == "cuda":
#         assert peak_mem >= mem_after >= mem_before



# import torch
# from safetensors import safe_open
# # from src.transformer_video import WanDiscreteVideoTransformer

# ckpt_path = "/mnt/Meissonic/model/diffusion_pytorch_model.safetensors"

# # 1) 按你想匹配 wan2.1 的超参实例化（这里写一份常用配置，务必与 ckpt 对齐）
# model = WanDiscreteVideoTransformer(
#     codebook_size=128,      # 离散侧自定义
#     vocab_size=129,
#     num_frames=2,
#     height=16,
#     width=16,
#     # Wan backbone 超参需与 ckpt 完全一致
#     model_type="t2v",
#     patch_size=(1, 2, 2),
#     in_dim=16,
#     dim=1536,
#     ffn_dim=8960,
#     freq_dim=256,
#     text_dim=4096,
#     out_dim=16,
#     num_heads=12,
#     num_layers=30,
#     window_size=(-1, -1),
#     qk_norm=True,
#     cross_attn_norm=True,
#     eps=1e-6,
# )

# # 2) 读取 safetensors
# state_dict = {}
# with safe_open(ckpt_path, framework="pt", device="cpu") as f:
#     for k in f.keys():
#         state_dict[k] = f.get_tensor(k)

# # 3) 尝试加载到 backbone（不碰 token_embedding/logits_head）
# missing, unexpected = model.backbone.load_state_dict(state_dict, strict=False)

# print("Missing keys:", missing[:50], "... total", len(missing))
# print("Unexpected keys:", unexpected[:50], "... total", len(unexpected))
# print("Backbone params (M):", sum(p.numel() for p in model.backbone.parameters()) / 1e6)
# print("Params (M):", sum(p.numel() for p in model.parameters()) / 1e6)

# # if __name__ == '__main__':
# #     # test_wan_discrete_video_transformer_forward_and_shapes()
# #     print('WanDiscreteVideoTransformer forward pass test: PASSED')