models/wan/causal_model.py

from models.wan.wan_base.modules.attention import attention
from models.wan.wan_base.modules.model import (
    WanRMSNorm,
    rope_apply,
    WanLayerNorm,
    WAN_CROSSATTENTION_CLASSES,
    Head,
    rope_params,
    MLPProj,
    sinusoidal_embedding_1d
)
from torch.nn.attention.flex_attention import create_block_mask, flex_attention
from diffusers.configuration_utils import ConfigMixin, register_to_config
from torch.nn.attention.flex_attention import BlockMask
from diffusers.models.modeling_utils import ModelMixin
import torch.nn.functional as F
import torch.nn as nn
import torch
import math
from collections import OrderedDict
import torch.distributed as dist
import warnings

try:
    import flash_attn_interface
    FLASH_ATTN_AVAILABLE = True
except (ImportError, ModuleNotFoundError):
    try:
        from flash_attn import flash_attn_interface
        FLASH_ATTN_AVAILABLE = True
    except (ImportError, ModuleNotFoundError):
        flash_attn_interface = None
        FLASH_ATTN_AVAILABLE = False

# wan 1.3B model has a weird channel / head configurations and require max-autotune to work with flexattention
# see https://github.com/pytorch/pytorch/issues/133254
# change to default for other models
flex_attention = torch.compile(
    flex_attention, dynamic=False, mode="max-autotune")

_CAUSAL_ROPE_FREQ_CACHE = OrderedDict()
_CAUSAL_ROPE_FREQ_CACHE_SIZE = 16


def _causal_rope_cache_key(freqs, f, h, w, start_frame, device):
    return (
        freqs,
        device.type,
        device.index,
        f,
        h,
        w,
        start_frame,
    )


def _get_causal_rope_freqs(freqs_source, freqs_parts, f, h, w, start_frame, device):
    key = _causal_rope_cache_key(freqs_source, f, h, w, start_frame, device)
    cached = _CAUSAL_ROPE_FREQ_CACHE.get(key)
    if cached is not None:
        _CAUSAL_ROPE_FREQ_CACHE.move_to_end(key)
        return cached

    temporal, height, width = freqs_parts
    temporal_freqs = temporal[start_frame:start_frame + f].repeat_interleave(h * w, dim=0)
    height_freqs = height[:h].repeat_interleave(w, dim=0).repeat(f, 1)
    width_freqs = width[:w].repeat(h, 1).repeat(f, 1)
    rope_freqs = torch.cat([temporal_freqs, height_freqs, width_freqs], dim=-1).unsqueeze(1)

    _CAUSAL_ROPE_FREQ_CACHE[key] = rope_freqs
    if len(_CAUSAL_ROPE_FREQ_CACHE) > _CAUSAL_ROPE_FREQ_CACHE_SIZE:
        _CAUSAL_ROPE_FREQ_CACHE.popitem(last=False)
    return rope_freqs


def _prepare_causal_rope_cache(grid_sizes, freqs, start_frame=0):
    c = freqs.shape[1]
    freqs_parts = freqs.split([c - 2 * (c // 3), c // 3, c // 3], dim=1)

    if isinstance(start_frame, torch.Tensor):
        start_frames = start_frame.tolist()
    else:
        start_frames = [int(start_frame)] * grid_sizes.shape[0]

    rope_cache = []
    for grid_size, sf in zip(grid_sizes.tolist(), start_frames):
        f, h, w = grid_size
        seq_len = f * h * w
        rope_freqs = _get_causal_rope_freqs(freqs, freqs_parts, f, h, w, sf, freqs.device)
        rope_cache.append((seq_len, rope_freqs))
    return rope_cache


def causal_rope_apply(x, grid_sizes, freqs, start_frame=0, rope_cache=None):
    n = x.size(2)

    if rope_cache is None:
        rope_cache = _prepare_causal_rope_cache(grid_sizes, freqs, start_frame=start_frame)

    output = x.clone()

    for i, (seq_len, freqs_i) in enumerate(rope_cache):
        x_i = torch.view_as_complex(
            x[i, :seq_len].to(torch.float64).reshape(seq_len, n, -1, 2)
        )
        output[i, :seq_len] = torch.view_as_real(x_i * freqs_i).flatten(2).type_as(x)

    return output


def attention_with_kvcache_fallback(q, k_cache, v_cache, cache_seqlens):
    out_dtype = q.dtype
    max_seq_len = k_cache.shape[1]

    def prepare_inputs(q_tensor, k_tensor, v_tensor):
        if q_tensor.device.type == "cpu" and q_tensor.dtype in (torch.float16, torch.bfloat16):
            q_tensor = q_tensor.float()
            k_tensor = k_tensor.float()
            v_tensor = v_tensor.float()
        return q_tensor, k_tensor, v_tensor

    # Fast path: every sample uses the same fully valid cache span.
    if torch.all(cache_seqlens == max_seq_len):
        q_all = q.transpose(1, 2)
        k_all = k_cache.transpose(1, 2)
        v_all = v_cache.transpose(1, 2)
        q_all, k_all, v_all = prepare_inputs(q_all, k_all, v_all)
        x = F.scaled_dot_product_attention(
            q_all,
            k_all,
            v_all,
            attn_mask=None,
            dropout_p=0.0,
            # Keep parity with flash_attn_with_kvcache(..., causal=False).
            is_causal=False,
        )
        return x.transpose(1, 2).to(out_dtype).contiguous()

    outputs = []
    for batch_idx, seq_len in enumerate(cache_seqlens.tolist()):
        q_i = q[batch_idx:batch_idx + 1].transpose(1, 2)
        k_i = k_cache[batch_idx:batch_idx + 1, :seq_len].transpose(1, 2)
        v_i = v_cache[batch_idx:batch_idx + 1, :seq_len].transpose(1, 2)
        q_i, k_i, v_i = prepare_inputs(q_i, k_i, v_i)

        x_i = F.scaled_dot_product_attention(
            q_i,
            k_i,
            v_i,
            attn_mask=None,
            dropout_p=0.0,
            # Keep parity with flash_attn_with_kvcache(..., causal=False).
            is_causal=False,
        )
        outputs.append(x_i.transpose(1, 2).to(out_dtype))

    return torch.cat(outputs, dim=0).contiguous()


class CausalWanSelfAttention(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
        
        self.sink_size = 3
        self.adapt_sink_thr = -1
        self.evict_idx = None

        # 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,
        block_mask,
        kv_cache=None,
        current_start=0,
        current_end=0,
        causal_rope_cache=None,
    ):
        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]
            block_mask (BlockMask)
        """
        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)

        if kv_cache is None:
            roped_query = rope_apply(q, grid_sizes, freqs).type_as(v)
            roped_key = rope_apply(k, grid_sizes, freqs).type_as(v)

            padded_length = math.ceil(q.shape[1] / 128) * 128 - q.shape[1]
            padded_roped_query = torch.cat(
                [roped_query,
                 torch.zeros([q.shape[0], padded_length, q.shape[2], q.shape[3]],
                             device=q.device, dtype=v.dtype)],
                dim=1
            )

            padded_roped_key = torch.cat(
                [roped_key, torch.zeros([k.shape[0], padded_length, k.shape[2], k.shape[3]],
                                        device=k.device, dtype=v.dtype)],
                dim=1
            )

            padded_v = torch.cat(
                [v, torch.zeros([v.shape[0], padded_length, v.shape[2], v.shape[3]],
                                device=v.device, dtype=v.dtype)],
                dim=1
            )

            x = flex_attention(
                query=padded_roped_query.transpose(2, 1),
                key=padded_roped_key.transpose(2, 1),
                value=padded_v.transpose(2, 1),
                block_mask=block_mask
            )[:, :, :-padded_length].transpose(2, 1)
        else:
            frame_seqlen = math.prod(grid_sizes[0][1:]).item()
            current_start_frame = current_start // frame_seqlen
            roped_query = causal_rope_apply(
                q,
                grid_sizes,
                freqs,
                start_frame=current_start_frame,
                rope_cache=causal_rope_cache,
            ).type_as(v)
            roped_key = causal_rope_apply(
                k,
                grid_sizes,
                freqs,
                start_frame=current_start_frame,
                rope_cache=causal_rope_cache,
            ).type_as(v)

            seq_lens = []
            kv_cache_size = kv_cache["k"].shape[1]
            cache_bs = kv_cache['k'].shape[0]
            
            # Ring-buffer queue init
            if self.evict_idx is None:
                self.evict_idx = [[]]

            if len(self.evict_idx) < cache_bs:
                self.evict_idx = [self.evict_idx[0].copy() for _ in range(cache_bs)]
                
            for i, c_start in enumerate(current_start):
                num_new_tokens = roped_query.shape[1]
                current_end = c_start + roped_query.shape[1]
                sink_tokens = self.sink_size * frame_seqlen
                
                if sink_tokens > 0 and self.adapt_sink_thr > -1 and v.shape[1] <= frame_seqlen:
                    # Caculate similarity between new keys/values and the oldest ones in the cache
                    k_sink_mean = kv_cache["k"][i:i+1, :sink_tokens].reshape(self.sink_size, frame_seqlen, -1).mean(1)
                    k_new_mean = roped_key[i:i+1].reshape(1, frame_seqlen, -1).mean(1)
                    k_cos_sim = torch.cosine_similarity(k_sink_mean, k_new_mean, dim=-1)

                    v_sink_mean = kv_cache["v"][i:i+1, :sink_tokens].reshape(self.sink_size, frame_seqlen, -1).mean(1)
                    v_new_mean = v[i:i+1].reshape(1, frame_seqlen, -1).mean(1)
                    v_cos_sim = torch.cosine_similarity(v_sink_mean, v_new_mean, dim=-1)

                    avg_cos_sim = (k_cos_sim + v_cos_sim)/2
                    # When the similarity is low, refresh the sink
                    if avg_cos_sim.min() < self.adapt_sink_thr:
                        idx = torch.argmin(avg_cos_sim).item()
                        temp_evict_idx = (idx+1) * frame_seqlen
                        self.evict_idx[i].insert(0, temp_evict_idx)

                # If we are using local attention and the current KV cache size is larger than the local attention size, we need to truncate the KV cache
                if current_end > kv_cache_size or kv_cache["local_end_index"][i]>=kv_cache_size:
                    kv_cache["global_end_index"][i].fill_(c_start)
                    kv_cache["local_end_index"][i].fill_(kv_cache_size)

                    target_end = self.evict_idx[i][0]

                    # current_step = kv_cache['current_step']
                    # Update the buffer
                    if cache_bs==1 and kv_cache['current_step'] > 1:
                        kv_cache['current_step']-=1
                    else:
                        evict_idx = self.evict_idx[i].pop(0)
                        if evict_idx > sink_tokens:
                            self.evict_idx[i].append(evict_idx)
                        kv_cache['current_step']=kv_cache['total_steps']

                    # print(f"self.evict_idx: {self.evict_idx[i]}, total steps: {kv_cache['total_steps']}, current step: {current_step}, target: {target_end-num_new_tokens}:{target_end}, kv size:{kv_cache_size}")

                    # Newly added cache covers the oldest one
                    kv_cache["k"][i:i+1, target_end-num_new_tokens:target_end] = roped_key[i:i+1]
                    kv_cache["v"][i:i+1, target_end-num_new_tokens:target_end] = v[i:i+1]

                    local_end_index = kv_cache["local_end_index"][i].item()

                else:
                    local_end_index = kv_cache["local_end_index"][i].item() + current_end - kv_cache["global_end_index"][i].item()

                    rolling_end = (current_end + num_new_tokens).item()
                    if rolling_end > self.sink_size * frame_seqlen and rolling_end <= kv_cache_size \
                        and (not self.evict_idx[i] or self.evict_idx[i][-1] != rolling_end):
                        self.evict_idx[i].append(rolling_end)

                    local_start_index = local_end_index - num_new_tokens
                    # print(f"target: {local_start_index}:{local_end_index}")
                    kv_cache["k"][i:i+1, local_start_index:local_end_index] = roped_key[i:i+1]
                    kv_cache["v"][i:i+1, local_start_index:local_end_index] = v[i:i+1]

                seq_lens.append(local_end_index)

                kv_cache["global_end_index"][i].fill_(current_end)
                kv_cache["local_end_index"][i].fill_(local_end_index)
            
            seq_lens = torch.tensor(seq_lens, dtype=torch.int32, device=roped_query.device)

            max_seq_len = int(seq_lens.max().item())
            k_cache = kv_cache["k"][:, :max_seq_len]
            v_cache = kv_cache["v"][:, :max_seq_len]

            if FLASH_ATTN_AVAILABLE:
                try:
                    with torch.cuda.device(roped_query.device):
                        x = flash_attn_interface.flash_attn_with_kvcache(
                            q=roped_query,
                            k_cache=k_cache,
                            v_cache=v_cache,
                            cache_seqlens=seq_lens,
                        )
                except RuntimeError as exc:
                    if "DeviceType::CUDA" not in str(exc):
                        raise
                    warnings.warn(
                        "flash_attn_with_kvcache failed on the current GPU; "
                        "falling back to scaled_dot_product_attention.",
                        stacklevel=2,
                    )
                    x = attention_with_kvcache_fallback(
                        q=roped_query,
                        k_cache=k_cache,
                        v_cache=v_cache,
                        cache_seqlens=seq_lens,
                    )
            else:
                warnings.warn(
                    "flash_attn is not installed; falling back to "
                    "scaled_dot_product_attention for KV-cache attention.",
                    stacklevel=2,
                )
                x = attention_with_kvcache_fallback(
                    q=roped_query,
                    k_cache=k_cache,
                    v_cache=v_cache,
                    cache_seqlens=seq_lens,
                )

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


class CausalWanAttentionBlock(nn.Module):

    def __init__(self,
                 cross_attn_type,
                 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 = CausalWanSelfAttention(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 = WAN_CROSSATTENTION_CLASSES[cross_attn_type](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,
        block_mask,
        kv_cache=None,
        crossattn_cache=None,
        current_start=0,
        current_end=0,
        causal_rope_cache=None,
    ):
        r"""
        Args:
            x(Tensor): Shape [B, L, C]
            e(Tensor): Shape [B, F, 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]
        """
        num_frames, frame_seqlen = e.shape[1], x.shape[1] // e.shape[1]
        # assert e.dtype == torch.float32
        # with amp.autocast(dtype=torch.float32):
        e = (self.modulation.unsqueeze(1) + e).chunk(6, dim=2)
        # assert e[0].dtype == torch.float32

        # self-attention
        y = self.self_attn(
            (self.norm1(x).unflatten(dim=1, sizes=(num_frames, frame_seqlen))
             * (1 + e[1]) + e[0]).flatten(1, 2),
            seq_lens, grid_sizes,
            freqs, block_mask, kv_cache, current_start, current_end, causal_rope_cache)

        # with amp.autocast(dtype=torch.float32):
        x = x + (y.unflatten(dim=1, sizes=(num_frames, frame_seqlen))
                 * e[2]).flatten(1, 2)

        # cross-attention & ffn function
        def cross_attn_ffn(x, context, context_lens, e, crossattn_cache=None):
            x = x + self.cross_attn(self.norm3(x), context,
                                    context_lens, crossattn_cache=crossattn_cache)
            y = self.ffn(
                (self.norm2(x).unflatten(dim=1, sizes=(num_frames,
                 frame_seqlen)) * (1 + e[4]) + e[3]).flatten(1, 2)
            )
            # with amp.autocast(dtype=torch.float32):
            x = x + (y.unflatten(dim=1, sizes=(num_frames,
                     frame_seqlen)) * e[5]).flatten(1, 2)
            return x

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


class CausalHead(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, F, 1, C]
        """
        # assert e.dtype == torch.float32
        # with amp.autocast(dtype=torch.float32):
        num_frames, frame_seqlen = e.shape[1], x.shape[1] // e.shape[1]
        e = (self.modulation.unsqueeze(1) + e).chunk(2, dim=2)
        x = (self.head(
            self.norm(x).unflatten(dim=1, sizes=(num_frames, frame_seqlen)) *
            (1 + e[1]) + e[0]))
        return x


class CausalWanModel(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']
    _supports_gradient_checkpointing = True

    @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']
        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
        cross_attn_type = 't2v_cross_attn' if model_type == 't2v' else 'i2v_cross_attn'
        self.blocks = nn.ModuleList([
            CausalWanAttentionBlock(cross_attn_type, dim, ffn_dim, num_heads,
                                    window_size, qk_norm, cross_attn_norm, eps)
            for _ in range(num_layers)
        ])

        # head
        self.head = CausalHead(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)

        if model_type == 'i2v':
            self.img_emb = MLPProj(1280, dim)

        # initialize weights
        self.init_weights()

        self.gradient_checkpointing = False

        self.block_mask = None

        self.num_frame_per_block = 1

    def _set_gradient_checkpointing(self, module, value=False):
        self.gradient_checkpointing = value

    @staticmethod
    def _prepare_blockwise_causal_attn_mask(
        device: torch.device | str, num_frames: int = 21,
        frame_seqlen: int = 1560, num_frame_per_block=1
    ) -> BlockMask:
        """
        we will divide the token sequence into the following format
        [1 latent frame] [1 latent frame] ... [1 latent frame]
        We use flexattention to construct the attention mask
        """
        total_length = num_frames * frame_seqlen

        # we do right padding to get to a multiple of 128
        padded_length = math.ceil(total_length / 128) * 128 - total_length

        ends = torch.zeros(total_length + padded_length,
                           device=device, dtype=torch.long)

        # Block-wise causal mask will attend to all elements that are before the end of the current chunk
        frame_indices = torch.arange(
            start=0,
            end=total_length,
            step=frame_seqlen * num_frame_per_block,
            device=device
        )

        for tmp in frame_indices:
            ends[tmp:tmp + frame_seqlen * num_frame_per_block] = tmp + \
                frame_seqlen * num_frame_per_block

        def attention_mask(b, h, q_idx, kv_idx):
            return (kv_idx < ends[q_idx]) | (q_idx == kv_idx)
            # return ((kv_idx < total_length) & (q_idx < total_length))  | (q_idx == kv_idx) # bidirectional mask

        block_mask = create_block_mask(attention_mask, B=None, H=None, Q_LEN=total_length + padded_length,
                                       KV_LEN=total_length + padded_length, _compile=False, device=device)

        import torch.distributed as dist
        if not dist.is_initialized() or dist.get_rank() == 0:
            print(
                f" cache a block wise causal mask with block size of {num_frame_per_block} frames")
            print(block_mask)

        return block_mask

    def _forward_inference(
        self,
        x,
        t,
        context,
        seq_len,
        clip_fea=None,
        y=None,
        kv_cache: dict = None,
        crossattn_cache: dict = None,
        current_start: int = 0,
        current_end: int = 0,
        block_mode: str = 'input',
        block_num: int = [-1],
        patched_x_shape: torch.Tensor = None,
    ):
        r"""
        Run the diffusion model with kv caching.
        See Algorithm 2 of CausVid paper https://arxiv.org/abs/2412.07772 for details.
        This function will be run for num_frame times.
        Process the latent frames one by one (1560 tokens each)

        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
        # params
        device = self.patch_embedding.weight.device
        if self.freqs.device != device:
            self.freqs = self.freqs.to(device)

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

            # embeddings
            x = [self.patch_embedding(u.unsqueeze(0)) for u in x]
            bsz, cch, tlen, hh, ww = x[0].shape
            patched_x_shape = torch.tensor([bsz, cch, tlen, hh, ww], dtype=torch.int64, device=device)
        else:
            bsz, cch, tlen, hh, ww = [int(i) for i in patched_x_shape.tolist()]
            x = [u.permute(1,0).reshape(bsz, cch, tlen, hh, ww) 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(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
        # with amp.autocast(dtype=torch.float32):
        e = self.time_embedding(
            sinusoidal_embedding_1d(self.freq_dim, t.flatten()).type_as(x))
        e0 = self.time_projection(e).unflatten(
            1, (6, self.dim)).unflatten(dim=0, sizes=t.shape)
        # 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:
            context_clip = self.img_emb(clip_fea)  # bs x 257 x dim
            context = torch.concat([context_clip, context], dim=1)

        # arguments
        kwargs = dict(
            e=e0,
            seq_lens=seq_lens,
            grid_sizes=grid_sizes,
            freqs=self.freqs,
            context=context,
            context_lens=context_lens,
            block_mask=self.block_mask
        )
        if kv_cache is not None:
            kwargs["causal_rope_cache"] = _prepare_causal_rope_cache(
                grid_sizes,
                self.freqs,
                start_frame=current_start // math.prod(grid_sizes[0][1:]).item(),
            )

        def create_custom_forward(module):
            def custom_forward(*inputs, **kwargs):
                return module(*inputs, **kwargs)
            return custom_forward
        
        for block_index, block in enumerate(self.blocks):
            if torch.is_grad_enabled() and self.gradient_checkpointing:
                assert False
            else:
                if (block_mode == 'output' or block_mode == 'middle') and block_index < block_num[0]:
                    continue
                if (block_mode == 'input' or block_mode == 'middle') and block_index == block_num[-1]:
                    return x, patched_x_shape
                kwargs.update(
                    {
                        "kv_cache": kv_cache[block_index],
                        "crossattn_cache": crossattn_cache[block_index],
                        "current_start": current_start,
                        "current_end": current_end
                    }
                )
                x = block(x, **kwargs)
        if block_mode == 'input' and block_num[-1] == len(self.blocks):
            return x, patched_x_shape

        # head
        x = self.head(x, e.unflatten(dim=0, sizes=t.shape).unsqueeze(2))

        # unpatchify
        x = self.unpatchify(x, grid_sizes)
        return torch.stack(x)

    def _forward_train(
        self,
        x,
        t,
        context,
        seq_len,
        clip_fea=None,
        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
            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
        # params
        device = self.patch_embedding.weight.device
        if self.freqs.device != device:
            self.freqs = self.freqs.to(device)

        # Construct blockwise causal attn mask
        if self.block_mask is None:
            self.block_mask = self._prepare_blockwise_causal_attn_mask(
                device, num_frames=x.shape[2],
                frame_seqlen=x.shape[-2] *
                x.shape[-1] // (self.patch_size[1] * self.patch_size[2]),
                num_frame_per_block=self.num_frame_per_block
            )

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

        # embeddings
        x = [self.patch_embedding(u.unsqueeze(0)) 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
        # with amp.autocast(dtype=torch.float32):
        e = self.time_embedding(
            sinusoidal_embedding_1d(self.freq_dim, t.flatten()).type_as(x))
        e0 = self.time_projection(e).unflatten(
            1, (6, self.dim)).unflatten(dim=0, sizes=t.shape)
        # 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:
            context_clip = self.img_emb(clip_fea)  # bs x 257 x dim
            context = torch.concat([context_clip, context], dim=1)

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

        def create_custom_forward(module):
            def custom_forward(*inputs, **kwargs):
                return module(*inputs, **kwargs)
            return custom_forward

        for block_index, block in enumerate(self.blocks):
            if torch.is_grad_enabled() and self.gradient_checkpointing:
                x = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(block),
                    x, **kwargs,
                    use_reentrant=False,
                )
            else:
                x = block(x, **kwargs)

        # head
        x = self.head(x, e.unflatten(dim=0, sizes=t.shape).unsqueeze(2))

        # unpatchify
        x = self.unpatchify(x, grid_sizes)
        return torch.stack(x)

    def forward(
        self,
        *args,
        **kwargs
    ):
        if kwargs.get('kv_cache', None) is not None:
            return self._forward_inference(*args, **kwargs)
        else:
            return self._forward_train(*args, **kwargs)

    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)