example.py

from creativelib.modules.attn import AttnConfig, create_attn_from_cfg, AttnRuntime, create_attn_runtime

from creativelib.modules.core import SparseAttnConfig
import torch 
import torch.nn as nn
from creativelib.utils.misc import align_up, div_up
from typing import Optional
import math 
from creativelib.attn.rope import rope_apply_fused

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
def sinusoidal_embedding_1d(dim, position):
    # preprocess
    assert dim % 2 == 0
    half = dim // 2
    position = position.type(torch.float64)

    # calculation
    sinusoid = torch.outer(
        position, torch.pow(10000, -torch.arange(half).to(position).div(half)))
    x = torch.cat([torch.cos(sinusoid), torch.sin(sinusoid)], dim=1)
    return x
def pad_freqs(original_tensor, target_len):
    seq_len, s1, s2 = original_tensor.shape
    pad_size = target_len - seq_len
    padding_tensor = torch.ones(
        pad_size,
        s1,
        s2,
        dtype=original_tensor.dtype,
        device=original_tensor.device)
    padded_tensor = torch.cat([original_tensor, padding_tensor], dim=0)
    return padded_tensor

def get_single_freq(freq_base, seq_length_after_cp, f, h, w, head_size: int = 128, reorder_inds: Optional[torch.Tensor] = None):
    seq_len = f * h * w
    c = head_size // 2
    freq_base = freq_base.split([c - 2 * (c // 3), c // 3, c // 3], dim=1)
    freqs_i = torch.cat([
        freq_base[0][:f].view(f, 1, 1, -1).expand(f, h, w, -1),
        freq_base[1][:h].view(1, h, 1, -1).expand(f, h, w, -1),
        freq_base[2][:w].view(1, 1, w, -1).expand(f, h, w, -1)
    ],
            dim=-1).reshape(seq_len, 1, -1)
    sp_size = 1
    sp_rank = 0
    if reorder_inds is not None:
        freqs_i = freqs_i[reorder_inds]
    # apply rotary embedding
    freqs_i = pad_freqs(freqs_i, seq_length_after_cp * sp_size)
    s_per_rank = seq_length_after_cp
    freqs_i_rank = freqs_i[(sp_rank * s_per_rank):((sp_rank + 1) *
                                                    s_per_rank), :, :]
    return freqs_i_rank.view(torch.float64).to(torch.float32)

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))

    @torch.compile(fullgraph=True, dynamic=True)
    def forward(self, x):
        r"""
        Args:
            x(Tensor): Shape [B, L, C]
        """
        return self._norm(x.float()).type_as(x) * self.weight

    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]
        """
        return super().forward(x.float()).type_as(x)


class WanSelfAttention(nn.Module):

    def __init__(self,
                 dim,
                 num_heads,
                 window_size=(-1, -1),
                 qk_norm=True,
                 eps=1e-6,
                 attn_cfg: Optional[AttnConfig] = None):
        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()
        if attn_cfg is None:
            attn_cfg = AttnConfig("sdpa")
        self.attn = create_attn_from_cfg(attn_cfg, num_heads, self.head_dim)

    def forward(self, x, freqs, attn_runtime):
        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 q_fn(x):
            q = self.norm_q(self.q(x)).view(b, s, n, d)
            q = rope_apply_fused(q, freqs)
            return q
        def k_fn(x):
            k = self.norm_k(self.k(x)).view(b, s, n, d)
            k = rope_apply_fused(k, freqs)
            return k
        def v_fn(x):
            v = self.v(x).view(b, s, n, d)
            return v
        x = self.attn(
            q_fn,
            k_fn,
            v_fn,
            attn_runtime=attn_runtime,
            x=x)
        # 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,
                 self_attn_cfg: Optional[AttnConfig] = None):
        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_attn_cfg)
        self.norm3 = WanLayerNorm(
            dim, eps,
            elementwise_affine=True) if cross_attn_norm else nn.Identity()
        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)
    
    @torch.compile(dynamic=True)
    def forward(
        self,
        x,
        e,
        seq_lens,
        grid_sizes,
        freqs,
        context,
        context_lens,
        attn_runtime: AttnRuntime,
    ):
        r"""
        Args:
            x(Tensor): Shape [B, L, C]
            e(Tensor): Shape [B, 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]
        """
        assert e.dtype == torch.float32
        with torch.amp.autocast("cuda", dtype=torch.float32):
            e = (self.modulation + e).chunk(6, dim=1)
        assert e[0].dtype == torch.float32

        # self-attention
        y = self.self_attn(
            self.norm1(x).float() * (1 + e[1]) + e[0], freqs, attn_runtime)

        with torch.amp.autocast("cuda", dtype=torch.float32):
            x = x + y * e[2]

        # cross-attention & ffn function
        def cross_attn_ffn(x, context, context_lens, e):
            # for simplicity, we remove cross attn here.
            x = x # + self.cross_attn(self.norm3(x), context, context_lens, is_uncond, context_nag, nag_params)
            y = self.ffn(self.norm2(x).float() * (1 + e[4]) + e[3])
            with torch.amp.autocast("cuda", dtype=torch.float32):
                x = x + y * e[5]
            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)

    @torch.compile
    def forward(self, x, e):
        r"""
        Args:
            x(Tensor): Shape [B, L1, C]
            e(Tensor): Shape [B, C]
        """
        assert e.dtype == torch.float32
        with torch.amp.autocast("cuda", dtype=torch.float32):
            e = (self.modulation + e.unsqueeze(1)).chunk(2, dim=1)
            x = (self.head(self.norm(x) * (1 + e[1]) + e[0]))
        return x

class WanModel(nn.Module):
    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']

    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,
                 non_sparse_self_attn_cfg: Optional[AttnConfig] = None,
                 self_attn_cfg: Optional[AttnConfig] = None,
                 num_non_sparse_layers: int = 0,
                 # if empty use num_non_sparse_layers
                 non_sparse_layer_inds:Optional[list[int]] = None,
                 eps=1e-6):

        super().__init__()

        assert model_type in ['t2v', 'i2v']
        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
        if self_attn_cfg is None:
            self_attn_cfg = AttnConfig("sdpa")
        if non_sparse_self_attn_cfg is None:
            non_sparse_self_attn_cfg = AttnConfig("sdpa")
        self._self_attn_cfg = self_attn_cfg
        self._non_sparse_self_attn_cfg = non_sparse_self_attn_cfg
        num_dense_self_block = num_non_sparse_layers
        # non_sparse_layer_inds = None
        if non_sparse_layer_inds is None:
            non_sparse_layer_inds = list(range(num_dense_self_block))
        else:
            assert non_sparse_layer_inds
            for i in range(len(non_sparse_layer_inds)):
                idx = non_sparse_layer_inds[i]
                if idx < 0:
                    non_sparse_layer_inds[i] += num_layers

        # 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
        attn_block_list: list[WanAttentionBlock] = []
        for j in range(num_layers):
            if j in non_sparse_layer_inds:
                self_attn_cfg_layer = non_sparse_self_attn_cfg
            else:
                self_attn_cfg_layer = self_attn_cfg
            attn_block_list.append(WanAttentionBlock(dim, ffn_dim, num_heads,
                            window_size, qk_norm, cross_attn_norm, eps,
                            self_attn_cfg_layer))
        self.blocks = nn.ModuleList(attn_block_list)
        # 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
        with torch.device("cpu"):
            # prevent outside meta init.
            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)

    
    def forward(
        self,
        x,
        t,
        context,
        seq_len,
        clip_fea=None,
        y=None,
        sparse_attn_cfg_override=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
            clip_fea (Tensor, *optional*):
                CLIP image features for image-to-video mode
            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 clip_fea is not None and y is not None
            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
        # run conv3d in f32 to avoid torch using slow native conv3d on 4090 + torch 2.9.1
        print([u.shape for u in x])
        x_tensor = torch.stack(x)
        # run conv3d in f32 to avoid torch using slow native conv3d
        with torch.amp.autocast("cuda", dtype=torch.float32):
            x_tensor = self.patch_embedding(x_tensor)
        grid_sizes = torch.stack(
            [torch.tensor(u.shape[1:], dtype=torch.long) for u in x_tensor])
        x_spatial_shapes = [x_tensor.shape[2:]]
        x_seqlens = [x_tensor.shape[2] * x_tensor.shape[3] * x_tensor.shape[4]]
        print(x_spatial_shapes, x_seqlens)
        sparse_cfg = self._self_attn_cfg
        if sparse_attn_cfg_override is not None:
            sparse_cfg = sparse_attn_cfg_override
        assert isinstance(sparse_cfg, AttnConfig)

        attn_rt = create_attn_runtime([sparse_cfg, self._non_sparse_self_attn_cfg], 
            x_spatial_shapes, x_seqlens, x_seqlens, x[0].device)
        x = x_tensor.flatten(2).transpose(1,2)

        # x = [u.flatten(2).transpose(1, 2) for u in x]
        seq_lens = torch.tensor([u.size(0) for u in x], dtype=torch.long)
        assert seq_lens.max() <= seq_len
        real_seqlen = x[0].size(0)
        x = torch.stack([
            torch.cat([u, u.new_zeros(seq_len - u.size(0), u.size(1))], dim=0)
            for u in x
        ])

        # time embeddings
        with torch.amp.autocast("cuda", dtype=torch.float32):
            e = self.time_embedding(
                sinusoidal_embedding_1d(self.freq_dim, t).float())
            e0 = self.time_projection(e).unflatten(1, (6, self.dim))
            assert e.dtype == torch.float32 and e0.dtype == torch.float32

        # context
        context_lens = None
        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
            ]))

        if clip_fea is not None and hasattr(self, 'img_emb'):
            context_clip = self.img_emb(clip_fea)  # bs x 257 x dim
            context = torch.concat([context_clip, context], dim=1)
        freqs_per_rank = get_single_freq(self.freqs, x.size(1), *grid_sizes[0].tolist(),
            )

        # arguments
        kwargs = dict(
            e=e0,
            seq_lens=seq_lens,
            grid_sizes=grid_sizes,
            freqs=freqs_per_rank,
            context=context,
            context_lens=context_lens,
            attn_runtime=attn_rt)

        for block_idx, block in enumerate(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 main_simple():
    sparse_cfg = SparseAttnConfig(
        tile_size=[4, 4, 4],
        enable_sta=False ,
        sta_kernel_size=[3, 3, 3],
        sta_separable=False,
        enable_standalone_sta=False,
        gate_per_head=False,
        gate_use_sigmoid=False,
        apply_gate_to_sparse=False,
        always_attn_first=False,
        coarse_attn_use_full_q=True,
        use_mpsa=True,
        use_coarse_bias=False,
        use_selected_qk_in_coarse=False,
        use_selected_qk_in_score=False,
        scale_coarse_by_rate=True,
        scale_range=[0.1, 4.0],
        sparsity_override=0.8,
        deterministic=False,
        sage_quantization=True,
        enable_tile=False,
        # coarse_scale_factor=4.0,
        enable_compact_tile=True,
        compact_tile_d_last=True,
    )
    use_sparse_sage_attn = True
    self_attn_cfg = AttnConfig(
        type="sparse_video_attn" if use_sparse_sage_attn else "sage_gluon_4090",
        vnsa_config=sparse_cfg,
    )
    dense_attn_cfg = AttnConfig(
        type="sage_gluon_4090",
    )
    num_heads = 40
    head_dim = 128
    attn = create_attn_from_cfg(self_attn_cfg, num_heads, head_dim)

    x_spatial_shapes = [[21, 52, 29]]
    x_seqlens = [21 * 52 * 29]
    x = torch.randn((1, x_seqlens[0], 5120), device="cuda", dtype=torch.bfloat16)
    # it's recommend to use dense attn for first two transformer block.
    attn_rt = create_attn_runtime([self_attn_cfg, dense_attn_cfg], 
        x_spatial_shapes, x_seqlens, x_seqlens, x.device)

    b, s, n, d = *x.shape[:2], num_heads, head_dim

    # functions is used for CP inference.
    def q_fn(x):
        q = x.view(b, s, n, d)
        return q
    def k_fn(x):
        k = x.view(b, s, n, d)
        return k
    def v_fn(x):
        v = x.view(b, s, n, d)
        return v

    x = attn(
        q_fn,
        k_fn,
        v_fn,
        attn_runtime=attn_rt,
        x=x)

def main_wan():
    sparse_cfg = SparseAttnConfig(
        tile_size=[4, 4, 4],
        enable_sta=False ,
        sta_kernel_size=[3, 3, 3],
        sta_separable=False,
        enable_standalone_sta=False,
        gate_per_head=False,
        gate_use_sigmoid=False,
        apply_gate_to_sparse=False,
        always_attn_first=False,
        coarse_attn_use_full_q=True,
        use_mpsa=True,
        use_coarse_bias=False,
        use_selected_qk_in_coarse=False,
        use_selected_qk_in_score=False,
        scale_coarse_by_rate=True,
        scale_range=[0.1, 4.0],
        sparsity_override=0.8,
        deterministic=False,
        sage_quantization=True,
        enable_tile=False,
        # coarse_scale_factor=4.0,
        enable_compact_tile=True,
        compact_tile_d_last=True,
    )
    use_sparse_sage_attn = True
    self_attn_cfg = AttnConfig(
        type="sparse_video_attn" if use_sparse_sage_attn else "sage_gluon_4090",
        vnsa_config=sparse_cfg,
    )
    dense_attn_cfg = AttnConfig(
        type="sage_gluon_4090",
    )

    model = WanModel(
        num_heads=16,
        num_layers=8,
        non_sparse_self_attn_cfg=dense_attn_cfg,
        self_attn_cfg=self_attn_cfg,
        num_non_sparse_layers=2, # it's recommend to use dense attn for first two transformer block.
    ).cuda().eval().bfloat16()
    x = torch.rand(16, 21, 52 * 2, 29 * 2).cuda()
    t = torch.randint(0, 1000, (1,)).cuda().float()
    # quantization requires padded seqlen to be multiple of 128
    padded_seqlen = align_up(21 * 52 * 29, 128)
    context = [torch.rand(77, 4096).cuda()]
    with torch.no_grad():
        with torch.amp.autocast(dtype=torch.bfloat16, device_type="cuda"):
            out = model([x], t, context, padded_seqlen)

if __name__ == "__main__":
    main_wan()