up

Co-authored-by: Yuyang Zhao <43061147+HeliosZhao@users.noreply.github.com>

scripts/convert_sana_video_to_diffusers.py

@@ -15,6 +15,7 @@ from diffusers import (
     AutoencoderKLWan,
     DPMSolverMultistepScheduler,
     FlowMatchEulerDiscreteScheduler,
+    SanaVideoCausalTransformer3DModel,
     SanaVideoPipeline,
     SanaVideoTransformer3DModel,
     UniPCMultistepScheduler,
@@ -24,7 +25,10 @@ from diffusers.utils.import_utils import is_accelerate_available
 
 CTX = init_empty_weights if is_accelerate_available else nullcontext
 
-ckpt_ids = ["Efficient-Large-Model/SANA-Video_2B_480p/checkpoints/SANA_Video_2B_480p.pth"]
+ckpt_ids = [
+    "Efficient-Large-Model/SANA-Video_2B_480p/checkpoints/SANA_Video_2B_480p.pth",
+    "Efficient-Large-Model/Sana-Video_2B_480p_LongLive/checkpoints/SANA_Video_2B_480p_LongLive.pth",
+]
 # https://github.com/NVlabs/Sana/blob/main/inference_video_scripts/inference_sana_video.py
 
 
@@ -98,6 +102,10 @@ def main(args):
     else:
         raise ValueError(f"Video size {args.video_size} is not supported.")
 
+    use_causal_linear_attn = False
+    if "Sana-Video_2B_480p_LongLive" in file_path:
+        use_causal_linear_attn = True
+
     for depth in range(layer_num):
         # Transformer blocks.
         converted_state_dict[f"transformer_blocks.{depth}.scale_shift_table"] = state_dict.pop(
@@ -201,7 +209,10 @@ def main(args):
             "rope_max_seq_len": 1024,
         }
 
-        transformer = SanaVideoTransformer3DModel(**transformer_kwargs)
+        if use_causal_linear_attn:
+            transformer = SanaVideoCausalTransformer3DModel(**transformer_kwargs)
+        else:
+            transformer = SanaVideoTransformer3DModel(**transformer_kwargs)
 
     transformer.load_state_dict(converted_state_dict, strict=True, assign=True)
 
@@ -314,7 +325,7 @@ if __name__ == "__main__":
         choices=["flow-dpm_solver", "flow-euler", "uni-pc"],
         help="Scheduler type to use.",
     )
-    parser.add_argument("--task", default="t2v", type=str, required=True, help="Task to convert, t2v or i2v.")
+    parser.add_argument("--task", default="t2v", type=str, required=True, choices=["t2v", "i2v"], help="Task to convert, t2v or i2v.")
     parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output pipeline.")
     parser.add_argument("--save_full_pipeline", action="store_true", help="save all the pipeline elements in one.")
     parser.add_argument("--dtype", default="fp32", type=str, choices=["fp32", "fp16", "bf16"], help="Weight dtype.")

src/diffusers/__init__.py

@@ -252,6 +252,7 @@ else:
             "SanaControlNetModel",
             "SanaTransformer2DModel",
             "SanaVideoTransformer3DModel",
+            "SanaVideoCausalTransformer3DModel",
             "SD3ControlNetModel",
             "SD3MultiControlNetModel",
             "SD3Transformer2DModel",
@@ -559,7 +560,7 @@ else:
             "SanaSprintImg2ImgPipeline",
             "SanaSprintPipeline",
             "SanaVideoPipeline",
-            "SanaVideoPipeline",
+            "LongSanaVideoPipeline",
             "SemanticStableDiffusionPipeline",
             "ShapEImg2ImgPipeline",
             "ShapEPipeline",
@@ -974,6 +975,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             QwenImageTransformer2DModel,
             SanaControlNetModel,
             SanaTransformer2DModel,
+            SanaVideoCausalTransformer3DModel,
             SanaVideoTransformer3DModel,
             SD3ControlNetModel,
             SD3MultiControlNetModel,
@@ -1215,6 +1217,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             LDMTextToImagePipeline,
             LEditsPPPipelineStableDiffusion,
             LEditsPPPipelineStableDiffusionXL,
+            LongSanaVideoPipeline,
             LTXConditionPipeline,
             LTXImageToVideoPipeline,
             LTXLatentUpsamplePipeline,

src/diffusers/models/__init__.py

@@ -108,6 +108,7 @@ if is_torch_available():
     _import_structure["transformers.transformer_prx"] = ["PRXTransformer2DModel"]
     _import_structure["transformers.transformer_qwenimage"] = ["QwenImageTransformer2DModel"]
     _import_structure["transformers.transformer_sana_video"] = ["SanaVideoTransformer3DModel"]
+    _import_structure["transformers.transformer_sana_video_causal"] = ["SanaVideoCausalTransformer3DModel"]
     _import_structure["transformers.transformer_sd3"] = ["SD3Transformer2DModel"]
     _import_structure["transformers.transformer_skyreels_v2"] = ["SkyReelsV2Transformer3DModel"]
     _import_structure["transformers.transformer_temporal"] = ["TransformerTemporalModel"]
@@ -217,6 +218,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             PRXTransformer2DModel,
             QwenImageTransformer2DModel,
             SanaTransformer2DModel,
+            SanaVideoCausalTransformer3DModel,
             SanaVideoTransformer3DModel,
             SD3Transformer2DModel,
             SkyReelsV2Transformer3DModel,

src/diffusers/models/transformers/__init__.py

@@ -40,6 +40,7 @@ if is_torch_available():
     from .transformer_prx import PRXTransformer2DModel
     from .transformer_qwenimage import QwenImageTransformer2DModel
     from .transformer_sana_video import SanaVideoTransformer3DModel
+    from .transformer_sana_video_causal import SanaVideoCausalTransformer3DModel
     from .transformer_sd3 import SD3Transformer2DModel
     from .transformer_skyreels_v2 import SkyReelsV2Transformer3DModel
     from .transformer_temporal import TransformerTemporalModel

src/diffusers/models/transformers/transformer_sana_video.py

@@ -94,6 +94,98 @@ class GLUMBTempConv(nn.Module):
         return hidden_states
 
 
+class CachedGLUMBConvTemp(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        expand_ratio: float = 4,
+        norm_type: Optional[str] = None,
+        residual_connection: bool = True,
+    ) -> None:
+        super().__init__()
+
+        hidden_channels = int(expand_ratio * in_channels)
+        self.norm_type = norm_type
+        self.residual_connection = residual_connection
+
+        self.nonlinearity = nn.SiLU()
+        self.conv_inverted = nn.Conv2d(in_channels, hidden_channels * 2, 1, 1, 0)
+        self.conv_depth = nn.Conv2d(hidden_channels * 2, hidden_channels * 2, 3, 1, 1, groups=hidden_channels * 2)
+        self.conv_point = nn.Conv2d(hidden_channels, out_channels, 1, 1, 0, bias=False)
+
+        self.norm = None
+        if norm_type == "rms_norm":
+            self.norm = RMSNorm(out_channels, eps=1e-5, elementwise_affine=True, bias=True)
+
+        self.conv_temp = nn.Conv2d(
+            out_channels, out_channels, kernel_size=(3, 1), stride=1, padding=(1, 0), bias=False
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        save_kv_cache: bool = False,
+        kv_cache: Optional[list] = None,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, list]]:
+        """
+        hidden_states: shape [B, T, H, W, C]
+        kv_cache: list, with kv_cache[0/1/2] for optional cached states (only kv_cache[2] is used here for temporal)
+        """
+
+        if self.residual_connection:
+            residual = hidden_states
+
+        batch_size, num_frames, height, width, num_channels = hidden_states.shape
+        hidden_states = hidden_states.view(batch_size * num_frames, height, width, num_channels).permute(0, 3, 1, 2)
+
+        hidden_states = self.conv_inverted(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.conv_depth(hidden_states)
+        hidden_states, gate = torch.chunk(hidden_states, 2, dim=1)
+        hidden_states = hidden_states * self.nonlinearity(gate)
+
+        hidden_states = self.conv_point(hidden_states)
+
+        # Temporal aggregation with kv_cache support
+        hidden_states_temporal = hidden_states.view(batch_size, num_frames, num_channels, height * width).permute(
+            0, 2, 1, 3
+        )
+
+        padding_size = self.conv_temp.kernel_size[0] // 2  # usually 1
+        hidden_states_temporal_in = hidden_states_temporal
+        padded_size = 0
+
+        # If using cache, prepend cached frames from last chunk along time axis (dim 2)
+        if kv_cache is not None:
+            if len(kv_cache) < 3:
+                kv_cache.extend([None] * (3 - len(kv_cache)))
+            if kv_cache[2] is not None:
+                hidden_states_temporal_in = torch.cat([kv_cache[2], hidden_states_temporal], dim=2)
+                padded_size = kv_cache[2].shape[2]
+            # Save last padding_size frames for next chunk
+            if save_kv_cache:
+                kv_cache[2] = hidden_states_temporal[:, :, -padding_size:, :].detach().clone()
+        else:
+            if save_kv_cache:
+                kv_cache = [None, None, hidden_states_temporal[:, :, -padding_size:, :].detach().clone()]
+
+        t_conv_out = self.conv_temp(hidden_states_temporal_in)[:, :, padded_size:]
+        hidden_states = hidden_states_temporal + t_conv_out
+        hidden_states = hidden_states.permute(0, 2, 3, 1).view(batch_size, num_frames, height, width, num_channels)
+
+        if self.norm_type == "rms_norm":
+            hidden_states = self.norm(hidden_states.movedim(1, -1)).movedim(-1, 1)
+
+        if self.residual_connection:
+            hidden_states = hidden_states + residual
+
+        if kv_cache is not None or save_kv_cache:
+            return hidden_states, kv_cache
+        return hidden_states
+
+
 class SanaLinearAttnProcessor3_0:
     r"""
     Processor for implementing scaled dot-product linear attention.
@@ -172,6 +264,121 @@ class SanaLinearAttnProcessor3_0:
         return hidden_states
 
 
+class SanaCausalLinearAttnProcessor1_0:
+    r"""
+    Processor for implementing causal linear attention with KV cache support.
+    Designed for autoregressive generation scenarios.
+    """
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        rotary_emb: Optional[torch.Tensor] = None,
+        save_kv_cache: bool = False,
+        kv_cache: Optional[list] = None,
+    ) -> torch.Tensor:
+        original_dtype = hidden_states.dtype
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+
+        # Project input to query, key, value
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        # Apply normalization
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # Reshape to multi-head format: B, N, C -> B, N, H, C
+        query = query.unflatten(2, (attn.heads, -1))
+        key = key.unflatten(2, (attn.heads, -1))
+        value = value.unflatten(2, (attn.heads, -1))
+
+        # Apply lightweight linear attention kernel (ReLU)
+        query = F.relu(query)
+        key = F.relu(key)
+
+        # Apply rotary position embeddings
+        if rotary_emb is not None:
+
+            def apply_rotary_emb(
+                hidden_states: torch.Tensor,
+                freqs_cos: torch.Tensor,
+                freqs_sin: torch.Tensor,
+            ):
+                x1, x2 = hidden_states.unflatten(-1, (-1, 2)).unbind(-1)
+                cos = freqs_cos[..., 0::2]
+                sin = freqs_sin[..., 1::2]
+                out = torch.empty_like(hidden_states)
+                out[..., 0::2] = x1 * cos - x2 * sin
+                out[..., 1::2] = x1 * sin + x2 * cos
+                return out.type_as(hidden_states)
+
+            query_rotate = apply_rotary_emb(query, *rotary_emb)
+            key_rotate = apply_rotary_emb(key, *rotary_emb)
+
+        # Permute to attention computation format: B, N, H, C -> B, H, C, N
+        query = query.permute(0, 2, 3, 1)
+        key = key.permute(0, 2, 3, 1)
+        query_rotate = query_rotate.permute(0, 2, 3, 1)
+        key_rotate = key_rotate.permute(0, 2, 3, 1)
+        value = value.permute(0, 2, 3, 1)
+
+        # Cast to float for numerical stability
+        query_rotate, key_rotate, value = query_rotate.float(), key_rotate.float(), value.float()
+
+        # Linear attention computation with KV cache support
+        # Compute key sum for normalization: sum over sequence dimension
+        k_sum = key.sum(dim=-1, keepdim=True).transpose(-2, -1)  # B, H, 1, C
+
+        # Compute value-key product: V @ K^T
+        scores = torch.matmul(value, key_rotate.transpose(-1, -2))  # B, H, C, C
+
+        # Handle KV cache for autoregressive generation
+        if kv_cache is not None:
+            cached_vk, cached_k_sum = kv_cache[0], kv_cache[1]
+
+            # Save current step's KV to cache if requested
+            if save_kv_cache:
+                kv_cache[0] = scores.detach().clone()
+                kv_cache[1] = k_sum.detach().clone()
+
+            # Accumulate with previous cached values
+            if cached_vk is not None and cached_k_sum is not None:
+                scores = scores + cached_vk
+                k_sum = k_sum + cached_k_sum
+
+        # Normalization factor: 1 / (K_sum @ Q + epsilon)
+        z = 1 / (k_sum @ query + 1e-15)
+
+        # Final attention output: (V @ K^T) @ Q
+        hidden_states = torch.matmul(scores, query_rotate)
+
+        # Apply normalization
+        hidden_states = hidden_states * z
+
+        # Reshape back: B, H, C, N -> B, N, C
+        hidden_states = hidden_states.flatten(1, 2).transpose(1, 2)
+        hidden_states = hidden_states.to(original_dtype)
+
+        # Output projection
+        hidden_states = attn.to_out[0](hidden_states)
+        hidden_states = attn.to_out[1](hidden_states)
+
+        # Return with cache if applicable
+        if kv_cache is not None:
+            return hidden_states, kv_cache
+
+        return hidden_states
+
+
 class WanRotaryPosEmbed(nn.Module):
     def __init__(
         self,
@@ -454,6 +661,146 @@ class SanaVideoTransformerBlock(nn.Module):
         return hidden_states
 
 
+class SanaVideoCausalTransformerBlock(nn.Module):
+    r"""
+    Transformer block with KV cache support for causal linear attention.
+    Used in LongSana for autoregressive generation.
+    """
+
+    def __init__(
+        self,
+        dim: int = 2240,
+        num_attention_heads: int = 20,
+        attention_head_dim: int = 112,
+        dropout: float = 0.0,
+        num_cross_attention_heads: Optional[int] = 20,
+        cross_attention_head_dim: Optional[int] = 112,
+        cross_attention_dim: Optional[int] = 2240,
+        attention_bias: bool = True,
+        norm_elementwise_affine: bool = False,
+        norm_eps: float = 1e-6,
+        attention_out_bias: bool = True,
+        mlp_ratio: float = 3.0,
+        qk_norm: Optional[str] = "rms_norm_across_heads",
+        rope_max_seq_len: int = 1024,
+        self_attn_processor: Optional[nn.Module] = None,
+        ffn_processor: Optional[nn.Module] = None,
+    ) -> None:
+        super().__init__()
+
+        # 1. Self Attention - must use causal linear attention
+        if self_attn_processor is None:
+            self_attn_processor = SanaCausalLinearAttnProcessor1_0()
+        self.norm1 = nn.LayerNorm(dim, elementwise_affine=False, eps=norm_eps)
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            kv_heads=num_attention_heads if qk_norm is not None else None,
+            qk_norm=qk_norm,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=None,
+            processor=self_attn_processor,
+        )
+
+        # 2. Cross Attention
+        if cross_attention_dim is not None:
+            self.norm2 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+            self.attn2 = Attention(
+                query_dim=dim,
+                qk_norm=qk_norm,
+                kv_heads=num_cross_attention_heads if qk_norm is not None else None,
+                cross_attention_dim=cross_attention_dim,
+                heads=num_cross_attention_heads,
+                dim_head=cross_attention_head_dim,
+                dropout=dropout,
+                bias=True,
+                out_bias=attention_out_bias,
+                processor=SanaAttnProcessor2_0(),
+            )
+
+        # 3. Feed-forward - must use cached conv
+        if ffn_processor is None:
+            ffn_processor = CachedGLUMBConvTemp
+        self.ff = ffn_processor(dim, dim, mlp_ratio, norm_type=None, residual_connection=False)
+
+        self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        frames: int = None,
+        height: int = None,
+        width: int = None,
+        rotary_emb: Optional[torch.Tensor] = None,
+        save_kv_cache: bool = False,
+        kv_cache: Optional[list] = None,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, list]]:
+        batch_size = hidden_states.shape[0]
+
+        # 1. Modulation
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+            self.scale_shift_table[None, None] + timestep.reshape(batch_size, timestep.shape[1], 6, -1)
+        ).unbind(dim=2)
+
+        # 2. Self Attention with KV cache
+        norm_hidden_states = self.norm1(hidden_states)
+        norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+        norm_hidden_states = norm_hidden_states.to(hidden_states.dtype)
+
+        # Causal linear attention always supports kv_cache
+        attn_result = self.attn1(
+            norm_hidden_states,
+            rotary_emb=rotary_emb,
+            save_kv_cache=save_kv_cache,
+            kv_cache=kv_cache,
+        )
+        if isinstance(attn_result, tuple):
+            attn_output, kv_cache = attn_result
+        else:
+            attn_output = attn_result
+
+        hidden_states = hidden_states + gate_msa * attn_output
+
+        # 3. Cross Attention (no cache)
+        if self.attn2 is not None:
+            attn_output = self.attn2(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward with KV cache
+        norm_hidden_states = self.norm2(hidden_states)
+        norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        norm_hidden_states = norm_hidden_states.unflatten(1, (frames, height, width))
+
+        # Cached conv always supports kv_cache
+        ff_result = self.ff(
+            norm_hidden_states,
+            save_kv_cache=save_kv_cache,
+            kv_cache=kv_cache,
+        )
+        if isinstance(ff_result, tuple):
+            ff_output, kv_cache = ff_result
+        else:
+            ff_output = ff_result
+
+        ff_output = ff_output.flatten(1, 3)
+        hidden_states = hidden_states + gate_mlp * ff_output
+
+        if kv_cache is not None or save_kv_cache:
+            return hidden_states, kv_cache
+        return hidden_states
+
+
 class SanaVideoTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, AttentionMixin):
     r"""
     A 3D Transformer model introduced in [Sana-Video](https://huggingface.co/papers/2509.24695) family of models.
@@ -703,3 +1050,271 @@ class SanaVideoTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, Fro
             return (output,)
 
         return Transformer2DModelOutput(sample=output)
+
+
+class SanaVideoCausalTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, AttentionMixin):
+    r"""
+    A 3D Transformer model with KV cache support for LongSana autoregressive generation.
+
+    This model extends Sana-Video with causal linear attention and cached convolutions
+    to enable efficient long video generation through chunked processing.
+
+    Args:
+        in_channels (`int`, defaults to `16`):
+            The number of channels in the input.
+        out_channels (`int`, *optional*, defaults to `16`):
+            The number of channels in the output.
+        num_attention_heads (`int`, defaults to `20`):
+            The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, defaults to `112`):
+            The number of channels in each head.
+        num_layers (`int`, defaults to `20`):
+            The number of layers of Transformer blocks to use.
+        num_cross_attention_heads (`int`, *optional*, defaults to `20`):
+            The number of heads to use for cross-attention.
+        cross_attention_head_dim (`int`, *optional*, defaults to `112`):
+            The number of channels in each head for cross-attention.
+        cross_attention_dim (`int`, *optional*, defaults to `2240`):
+            The number of channels in the cross-attention output.
+        caption_channels (`int`, defaults to `2304`):
+            The number of channels in the caption embeddings.
+        mlp_ratio (`float`, defaults to `2.5`):
+            The expansion ratio to use in the GLUMBConv layer.
+        dropout (`float`, defaults to `0.0`):
+            The dropout probability.
+        attention_bias (`bool`, defaults to `False`):
+            Whether to use bias in the attention layer.
+        sample_size (`int`, defaults to `32`):
+            The base size of the input latent.
+        patch_size (`int`, defaults to `1`):
+            The size of the patches to use in the patch embedding layer.
+        norm_elementwise_affine (`bool`, defaults to `False`):
+            Whether to use elementwise affinity in the normalization layer.
+        norm_eps (`float`, defaults to `1e-6`):
+            The epsilon value for the normalization layer.
+        qk_norm (`str`, *optional*, defaults to `None`):
+            The normalization to use for the query and key.
+    """
+
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["SanaVideoCausalTransformerBlock", "SanaModulatedNorm"]
+    _skip_layerwise_casting_patterns = ["patch_embedding", "norm"]
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 16,
+        out_channels: Optional[int] = 16,
+        num_attention_heads: int = 20,
+        attention_head_dim: int = 112,
+        num_layers: int = 20,
+        num_cross_attention_heads: Optional[int] = 20,
+        cross_attention_head_dim: Optional[int] = 112,
+        cross_attention_dim: Optional[int] = 2240,
+        caption_channels: int = 2304,
+        mlp_ratio: float = 2.5,
+        dropout: float = 0.0,
+        attention_bias: bool = False,
+        sample_size: int = 30,
+        patch_size: Tuple[int, int, int] = (1, 2, 2),
+        norm_elementwise_affine: bool = False,
+        norm_eps: float = 1e-6,
+        interpolation_scale: Optional[int] = None,
+        guidance_embeds: bool = False,
+        guidance_embeds_scale: float = 0.1,
+        qk_norm: Optional[str] = "rms_norm_across_heads",
+        rope_max_seq_len: int = 1024,
+    ) -> None:
+        super().__init__()
+
+        out_channels = out_channels or in_channels
+        inner_dim = num_attention_heads * attention_head_dim
+
+        # 1. Patch & position embedding
+        self.rope = WanRotaryPosEmbed(attention_head_dim, patch_size, rope_max_seq_len)
+        self.patch_embedding = nn.Conv3d(in_channels, inner_dim, kernel_size=patch_size, stride=patch_size)
+
+        # 2. Additional condition embeddings
+        if guidance_embeds:
+            self.time_embed = SanaCombinedTimestepGuidanceEmbeddings(inner_dim)
+        else:
+            self.time_embed = AdaLayerNormSingle(inner_dim)
+
+        self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
+        self.caption_norm = RMSNorm(inner_dim, eps=1e-5, elementwise_affine=True)
+
+        # 3. Transformer blocks - use causal versions
+        self.transformer_blocks = nn.ModuleList(
+            [
+                SanaVideoCausalTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    num_cross_attention_heads=num_cross_attention_heads,
+                    cross_attention_head_dim=cross_attention_head_dim,
+                    cross_attention_dim=cross_attention_dim,
+                    attention_bias=attention_bias,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                    mlp_ratio=mlp_ratio,
+                    qk_norm=qk_norm,
+                    self_attn_processor=SanaCausalLinearAttnProcessor1_0(),
+                    ffn_processor=CachedGLUMBConvTemp,
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        # 4. Output blocks
+        self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5)
+        self.norm_out = SanaModulatedNorm(inner_dim, elementwise_affine=False, eps=1e-6)
+        self.proj_out = nn.Linear(inner_dim, math.prod(patch_size) * out_channels)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        timestep: torch.Tensor,
+        guidance: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        controlnet_block_samples: Optional[Tuple[torch.Tensor]] = None,
+        save_kv_cache: bool = False,
+        kv_cache: Optional[list] = None,
+        return_dict: bool = True,
+    ) -> Union[Tuple[torch.Tensor, ...], Transformer2DModelOutput]:
+        if attention_kwargs is not None:
+            attention_kwargs = attention_kwargs.copy()
+            lora_scale = attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if attention_kwargs is not None and attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
+        if attention_mask is not None and attention_mask.ndim == 2:
+            attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 1. Input
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        p_t, p_h, p_w = self.config.patch_size
+        post_patch_num_frames = num_frames // p_t
+        post_patch_height = height // p_h
+        post_patch_width = width // p_w
+
+        rotary_emb = self.rope(hidden_states)
+
+        hidden_states = self.patch_embedding(hidden_states)
+        hidden_states = hidden_states.flatten(2).transpose(1, 2)
+
+        if guidance is not None:
+            timestep, embedded_timestep = self.time_embed(
+                timestep.flatten(), guidance=guidance, hidden_dtype=hidden_states.dtype
+            )
+        else:
+            timestep, embedded_timestep = self.time_embed(
+                timestep.flatten(), batch_size=batch_size, hidden_dtype=hidden_states.dtype
+            )
+
+        timestep = timestep.view(batch_size, -1, timestep.size(-1))
+        embedded_timestep = embedded_timestep.view(batch_size, -1, embedded_timestep.size(-1))
+
+        encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+        encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+
+        encoder_hidden_states = self.caption_norm(encoder_hidden_states)
+
+        # 2. Transformer blocks with KV cache
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+            # Note: gradient checkpointing doesn't support kv_cache (requires tuple return)
+            if kv_cache is not None:
+                logger.warning("KV cache is not supported with gradient checkpointing. Disabling KV cache.")
+                kv_cache = None
+
+            for index_block, block in enumerate(self.transformer_blocks):
+                hidden_states = self._gradient_checkpointing_func(
+                    block,
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    post_patch_num_frames,
+                    post_patch_height,
+                    post_patch_width,
+                    rotary_emb,
+                )
+                if controlnet_block_samples is not None and 0 < index_block <= len(controlnet_block_samples):
+                    hidden_states = hidden_states + controlnet_block_samples[index_block - 1]
+
+        else:
+            for index_block, block in enumerate(self.transformer_blocks):
+                # Get kv_cache for this block if available
+                block_kv_cache = kv_cache[index_block] if kv_cache is not None else None
+
+                block_result = block(
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    post_patch_num_frames,
+                    post_patch_height,
+                    post_patch_width,
+                    rotary_emb,
+                    save_kv_cache=save_kv_cache,
+                    kv_cache=block_kv_cache,
+                )
+
+                # Handle return value (could be tensor or tuple)
+                if isinstance(block_result, tuple):
+                    hidden_states, updated_kv_cache = block_result
+                    if kv_cache is not None:
+                        kv_cache[index_block] = updated_kv_cache
+                else:
+                    hidden_states = block_result
+
+                if controlnet_block_samples is not None and 0 < index_block <= len(controlnet_block_samples):
+                    hidden_states = hidden_states + controlnet_block_samples[index_block - 1]
+
+        # 3. Normalization
+        hidden_states = self.norm_out(hidden_states, embedded_timestep, self.scale_shift_table)
+
+        hidden_states = self.proj_out(hidden_states)
+
+        # 5. Unpatchify
+        hidden_states = hidden_states.reshape(
+            batch_size, post_patch_num_frames, post_patch_height, post_patch_width, p_t, p_h, p_w, -1
+        )
+        hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
+        output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            if kv_cache is not None or save_kv_cache:
+                return (output, kv_cache)
+            return (output,)
+
+        if kv_cache is not None or save_kv_cache:
+            return Transformer2DModelOutput(sample=output), kv_cache
+        return Transformer2DModelOutput(sample=output)

src/diffusers/models/transformers/transformer_sana_video_causal.py

@@ -0,0 +1,820 @@
+# Copyright 2025 The HuggingFace Team and SANA-Video Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import Any, Dict, Optional, Tuple, Union, List
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
+from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers, BaseOutput
+from ..attention import AttentionMixin
+from ..attention_dispatch import dispatch_attention_fn
+from ..attention_processor import Attention
+from ..embeddings import PixArtAlphaTextProjection, TimestepEmbedding, Timesteps, get_1d_rotary_pos_embed
+from ..modeling_utils import ModelMixin
+from ..normalization import AdaLayerNormSingle, RMSNorm
+
+from dataclasses import dataclass
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+
+@dataclass
+class SanaBlockKvCache:
+    vk: Optional[torch.Tensor] = None
+    k_sum: Optional[torch.Tensor] = None
+    temporal_cache: Optional[torch.Tensor] = None
+    _enable_save: bool = False
+
+    def disable_save(self):
+        self._enable_save = False
+
+    def enable_save(self):
+        self._enable_save = True
+
+    def maybe_save(
+        self, 
+        vk: Optional[torch.Tensor]=None, 
+        k_sum: Optional[torch.Tensor]=None, 
+        temporal_cache: Optional[torch.Tensor]=None,
+    ):
+        if not self._enable_save:
+            return
+
+        if vk is not None:
+            self.vk = vk.detach().clone()
+        if k_sum is not None:
+            self.k_sum = k_sum.detach().clone()
+        if temporal_cache is not None:
+            self.temporal_cache = temporal_cache.detach().clone()
+
+
+@dataclass
+class SanaVideoCausalTransformer3DModelOutput(BaseOutput):
+    """
+    The output of [`SanaVideoCausalTransformer3DModel`].
+
+    Args:
+        sample (`torch.Tensor` of shape `(batch_size, num_frames, height, width, num_channels)`):
+            The hidden states output conditioned on the `encoder_hidden_states` input.
+        kv_cache (`SanaKvCache`, *optional*):
+            The KV cache for the transformer blocks.
+    """
+
+    sample: "torch.Tensor"  # noqa: F821
+    kv_caches: Optional[List[SanaBlockKvCache]] = None
+
+
+class CachedGLUMBConvTemp(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        expand_ratio: float = 4,
+        norm_type: Optional[str] = None,
+        residual_connection: bool = True,
+    ) -> None:
+        super().__init__()
+
+        hidden_channels = int(expand_ratio * in_channels)
+        self.norm_type = norm_type
+        self.residual_connection = residual_connection
+
+        self.nonlinearity = nn.SiLU()
+        self.conv_inverted = nn.Conv2d(in_channels, hidden_channels * 2, 1, 1, 0)
+        self.conv_depth = nn.Conv2d(hidden_channels * 2, hidden_channels * 2, 3, 1, 1, groups=hidden_channels * 2)
+        self.conv_point = nn.Conv2d(hidden_channels, out_channels, 1, 1, 0, bias=False)
+
+        self.norm = None
+        if norm_type == "rms_norm":
+            self.norm = RMSNorm(out_channels, eps=1e-5, elementwise_affine=True, bias=True)
+
+        self.conv_temp = nn.Conv2d(
+            out_channels, out_channels, kernel_size=(3, 1), stride=1, padding=(1, 0), bias=False
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        kv_cache: Optional[SanaBlockKvCache] = None,
+    ) -> Tuple[torch.Tensor, Optional[SanaBlockKvCache]]:
+        """
+        hidden_states: shape [B, T, H, W, C]
+        kv_cache: SanaBlockKvCache, with optional cached states (only temporal_cache is used here for temporal)
+        """
+
+        if self.residual_connection:
+            residual = hidden_states
+
+        batch_size, num_frames, height, width, num_channels = hidden_states.shape
+        hidden_states = hidden_states.view(batch_size * num_frames, height, width, num_channels).permute(0, 3, 1, 2)
+
+        hidden_states = self.conv_inverted(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.conv_depth(hidden_states)
+        hidden_states, gate = torch.chunk(hidden_states, 2, dim=1)
+        hidden_states = hidden_states * self.nonlinearity(gate)
+
+        hidden_states = self.conv_point(hidden_states)
+
+        # Temporal aggregation with kv_cache support
+        hidden_states_temporal = hidden_states.view(batch_size, num_frames, num_channels, height * width).permute(
+            0, 2, 1, 3
+        )
+
+        padding_size = self.conv_temp.kernel_size[0] // 2  # usually 1
+        hidden_states_temporal_in = hidden_states_temporal
+        padded_size = 0
+
+        # If using cache, prepend cached frames from last chunk along time axis (dim 2)
+        if kv_cache is not None:
+            if kv_cache.temporal_cache is not None:
+                hidden_states_temporal_in = torch.cat([kv_cache.temporal_cache, hidden_states_temporal], dim=2)
+                padded_size = kv_cache.temporal_cache.shape[2]
+            # Save last padding_size frames for next chunk
+            kv_cache.maybe_save(
+                temporal_cache=hidden_states_temporal[:, :, -padding_size:, :],
+            )
+
+        t_conv_out = self.conv_temp(hidden_states_temporal_in)[:, :, padded_size:]
+        hidden_states = hidden_states_temporal + t_conv_out
+        hidden_states = hidden_states.permute(0, 2, 3, 1).view(batch_size, num_frames, height, width, num_channels)
+
+        if self.norm_type == "rms_norm":
+            hidden_states = self.norm(hidden_states.movedim(1, -1)).movedim(-1, 1)
+
+        if self.residual_connection:
+            hidden_states = hidden_states + residual
+
+        return hidden_states, kv_cache
+
+
+class SanaCausalLinearAttnProcessor1_0:
+    r"""
+    Processor for implementing causal linear attention with KV cache support.
+    Designed for autoregressive generation scenarios.
+    """
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        rotary_emb: Optional[torch.Tensor] = None,
+        kv_cache: Optional[SanaBlockKvCache] = None,
+    ) -> Tuple[torch.Tensor, Optional[SanaBlockKvCache]]:
+        original_dtype = hidden_states.dtype
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+
+        # Project input to query, key, value
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        # Apply normalization
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # Reshape to multi-head format: B, N, C -> B, N, H, C
+        query = query.unflatten(2, (attn.heads, -1))
+        key = key.unflatten(2, (attn.heads, -1))
+        value = value.unflatten(2, (attn.heads, -1))
+
+        # Apply lightweight linear attention kernel (ReLU)
+        query = F.relu(query)
+        key = F.relu(key)
+
+        # Apply rotary position embeddings
+        if rotary_emb is not None:
+
+            def apply_rotary_emb(
+                hidden_states: torch.Tensor,
+                freqs_cos: torch.Tensor,
+                freqs_sin: torch.Tensor,
+            ):
+                x1, x2 = hidden_states.unflatten(-1, (-1, 2)).unbind(-1)
+                cos = freqs_cos[..., 0::2]
+                sin = freqs_sin[..., 1::2]
+                out = torch.empty_like(hidden_states)
+                out[..., 0::2] = x1 * cos - x2 * sin
+                out[..., 1::2] = x1 * sin + x2 * cos
+                return out.type_as(hidden_states)
+
+            query_rotate = apply_rotary_emb(query, *rotary_emb)
+            key_rotate = apply_rotary_emb(key, *rotary_emb)
+
+        # Permute to attention computation format: B, N, H, C -> B, H, C, N
+        query = query.permute(0, 2, 3, 1)
+        key = key.permute(0, 2, 3, 1)
+        query_rotate = query_rotate.permute(0, 2, 3, 1)
+        key_rotate = key_rotate.permute(0, 2, 3, 1)
+        value = value.permute(0, 2, 3, 1)
+
+        # Cast to float for numerical stability
+        query_rotate, key_rotate, value = query_rotate.float(), key_rotate.float(), value.float()
+
+        # Linear attention computation with KV cache support
+        # Compute key sum for normalization: sum over sequence dimension
+        k_sum = key.sum(dim=-1, keepdim=True).transpose(-2, -1)  # B, H, 1, C
+
+        # Compute value-key product: V @ K^T
+        scores = torch.matmul(value, key_rotate.transpose(-1, -2))  # B, H, C, C
+
+        # Handle KV cache for autoregressive generation
+        if kv_cache is not None:
+            cached_vk, cached_k_sum = kv_cache.vk, kv_cache.k_sum
+            kv_cache.maybe_save(vk=scores, k_sum=k_sum)
+            if cached_vk is not None and cached_k_sum is not None:
+                scores = scores + cached_vk
+                k_sum = k_sum + cached_k_sum
+
+        # Normalization factor: 1 / (K_sum @ Q + epsilon)
+        z = 1 / (k_sum @ query + 1e-15)
+
+        # Final attention output: (V @ K^T) @ Q
+        hidden_states = torch.matmul(scores, query_rotate)
+
+        # Apply normalization
+        hidden_states = hidden_states * z
+
+        # Reshape back: B, H, C, N -> B, N, C
+        hidden_states = hidden_states.flatten(1, 2).transpose(1, 2)
+        hidden_states = hidden_states.to(original_dtype)
+
+        # Output projection
+        hidden_states = attn.to_out[0](hidden_states)
+        hidden_states = attn.to_out[1](hidden_states)
+
+        return hidden_states, kv_cache
+
+
+# Copied from transformers.transformer_sana_video.WanRotaryPosEmbed
+class WanRotaryPosEmbed(nn.Module):
+    def __init__(
+        self,
+        attention_head_dim: int,
+        patch_size: Tuple[int, int, int],
+        max_seq_len: int,
+        theta: float = 10000.0,
+    ):
+        super().__init__()
+
+        self.attention_head_dim = attention_head_dim
+        self.patch_size = patch_size
+        self.max_seq_len = max_seq_len
+
+        h_dim = w_dim = 2 * (attention_head_dim // 6)
+        t_dim = attention_head_dim - h_dim - w_dim
+
+        self.t_dim = t_dim
+        self.h_dim = h_dim
+        self.w_dim = w_dim
+
+        freqs_dtype = torch.float32 if torch.backends.mps.is_available() else torch.float64
+
+        freqs_cos = []
+        freqs_sin = []
+
+        for dim in [t_dim, h_dim, w_dim]:
+            freq_cos, freq_sin = get_1d_rotary_pos_embed(
+                dim,
+                max_seq_len,
+                theta,
+                use_real=True,
+                repeat_interleave_real=True,
+                freqs_dtype=freqs_dtype,
+            )
+            freqs_cos.append(freq_cos)
+            freqs_sin.append(freq_sin)
+
+        self.register_buffer("freqs_cos", torch.cat(freqs_cos, dim=1), persistent=False)
+        self.register_buffer("freqs_sin", torch.cat(freqs_sin, dim=1), persistent=False)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        p_t, p_h, p_w = self.patch_size
+        ppf, pph, ppw = num_frames // p_t, height // p_h, width // p_w
+
+        split_sizes = [self.t_dim, self.h_dim, self.w_dim]
+
+        freqs_cos = self.freqs_cos.split(split_sizes, dim=1)
+        freqs_sin = self.freqs_sin.split(split_sizes, dim=1)
+
+        freqs_cos_f = freqs_cos[0][:ppf].view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_cos_h = freqs_cos[1][:pph].view(1, pph, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_cos_w = freqs_cos[2][:ppw].view(1, 1, ppw, -1).expand(ppf, pph, ppw, -1)
+
+        freqs_sin_f = freqs_sin[0][:ppf].view(ppf, 1, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_sin_h = freqs_sin[1][:pph].view(1, pph, 1, -1).expand(ppf, pph, ppw, -1)
+        freqs_sin_w = freqs_sin[2][:ppw].view(1, 1, ppw, -1).expand(ppf, pph, ppw, -1)
+
+        freqs_cos = torch.cat([freqs_cos_f, freqs_cos_h, freqs_cos_w], dim=-1).reshape(1, ppf * pph * ppw, 1, -1)
+        freqs_sin = torch.cat([freqs_sin_f, freqs_sin_h, freqs_sin_w], dim=-1).reshape(1, ppf * pph * ppw, 1, -1)
+
+        return freqs_cos, freqs_sin
+
+
+# Copied from transformers.transformer_sana_video.SanaModulatedNorm
+class SanaModulatedNorm(nn.Module):
+    def __init__(self, dim: int, elementwise_affine: bool = False, eps: float = 1e-6):
+        super().__init__()
+        self.norm = nn.LayerNorm(dim, elementwise_affine=elementwise_affine, eps=eps)
+
+    def forward(
+        self, hidden_states: torch.Tensor, temb: torch.Tensor, scale_shift_table: torch.Tensor
+    ) -> torch.Tensor:
+        hidden_states = self.norm(hidden_states)
+        shift, scale = (scale_shift_table[None, None] + temb[:, :, None].to(scale_shift_table.device)).unbind(dim=2)
+        hidden_states = hidden_states * (1 + scale) + shift
+        return hidden_states
+
+
+# Copied from transformers.transformer_sana_video.SanaCombinedTimestepGuidanceEmbeddings
+class SanaCombinedTimestepGuidanceEmbeddings(nn.Module):
+    def __init__(self, embedding_dim):
+        super().__init__()
+        self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+
+        self.guidance_condition_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.guidance_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+
+        self.silu = nn.SiLU()
+        self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=True)
+
+    def forward(self, timestep: torch.Tensor, guidance: torch.Tensor = None, hidden_dtype: torch.dtype = None):
+        timesteps_proj = self.time_proj(timestep)
+        timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype))  # (N, D)
+
+        guidance_proj = self.guidance_condition_proj(guidance)
+        guidance_emb = self.guidance_embedder(guidance_proj.to(dtype=hidden_dtype))
+        conditioning = timesteps_emb + guidance_emb
+
+        return self.linear(self.silu(conditioning)), conditioning
+
+
+# Copied from transformers.transformer_sana_video.SanaAttnProcessor2_0
+class SanaAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    _attention_backend = None
+    _parallel_config = None
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("SanaAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim)
+        key = key.view(batch_size, -1, attn.heads, head_dim)
+        value = value.view(batch_size, -1, attn.heads, head_dim)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        hidden_states = dispatch_attention_fn(
+            query,
+            key,
+            value,
+            attn_mask=attention_mask,
+            dropout_p=0.0,
+            is_causal=False,
+            backend=self._attention_backend,
+            parallel_config=self._parallel_config,
+        )
+        hidden_states = hidden_states.flatten(2, 3)
+        hidden_states = hidden_states.type_as(query)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class SanaVideoCausalTransformerBlock(nn.Module):
+    r"""
+    Transformer block with KV cache support for causal linear attention.
+    Used in LongSana for autoregressive generation.
+    """
+
+    def __init__(
+        self,
+        dim: int = 2240,
+        num_attention_heads: int = 20,
+        attention_head_dim: int = 112,
+        dropout: float = 0.0,
+        num_cross_attention_heads: Optional[int] = 20,
+        cross_attention_head_dim: Optional[int] = 112,
+        cross_attention_dim: Optional[int] = 2240,
+        attention_bias: bool = True,
+        norm_elementwise_affine: bool = False,
+        norm_eps: float = 1e-6,
+        attention_out_bias: bool = True,
+        mlp_ratio: float = 3.0,
+        qk_norm: Optional[str] = "rms_norm_across_heads",
+        rope_max_seq_len: int = 1024,
+    ) -> None:
+        super().__init__()
+
+        # 1. Self Attention - must use causal linear attention
+        self.norm1 = nn.LayerNorm(dim, elementwise_affine=False, eps=norm_eps)
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            kv_heads=num_attention_heads if qk_norm is not None else None,
+            qk_norm=qk_norm,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=None,
+            processor=SanaCausalLinearAttnProcessor1_0(),
+        )
+
+        # 2. Cross Attention
+        if cross_attention_dim is not None:
+            self.norm2 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+            self.attn2 = Attention(
+                query_dim=dim,
+                qk_norm=qk_norm,
+                kv_heads=num_cross_attention_heads if qk_norm is not None else None,
+                cross_attention_dim=cross_attention_dim,
+                heads=num_cross_attention_heads,
+                dim_head=cross_attention_head_dim,
+                dropout=dropout,
+                bias=True,
+                out_bias=attention_out_bias,
+                processor=SanaAttnProcessor2_0(),
+            )
+
+        # 3. Feed-forward - must use cached conv
+        self.ff = CachedGLUMBConvTemp(dim, dim, mlp_ratio, norm_type=None, residual_connection=False)
+
+        self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        frames: int = None,
+        height: int = None,
+        width: int = None,
+        rotary_emb: Optional[torch.Tensor] = None,
+        kv_cache: Optional[SanaBlockKvCache] = None,
+    ) -> Tuple[torch.Tensor, Optional[SanaBlockKvCache]]:
+        batch_size = hidden_states.shape[0]
+
+        # 1. Modulation
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+            self.scale_shift_table[None, None] + timestep.reshape(batch_size, timestep.shape[1], 6, -1)
+        ).unbind(dim=2)
+
+        # 2. Self Attention with KV cache
+        norm_hidden_states = self.norm1(hidden_states)
+        norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+        norm_hidden_states = norm_hidden_states.to(hidden_states.dtype)
+
+        # Causal linear attention always supports kv_cache
+        attn_output, kv_cache = self.attn1(
+            norm_hidden_states,
+            rotary_emb=rotary_emb,
+            kv_cache=kv_cache,
+        )
+        hidden_states = hidden_states + gate_msa * attn_output
+
+        # 3. Cross Attention (no cache)
+        if self.attn2 is not None:
+            attn_output = self.attn2(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward with KV cache
+        norm_hidden_states = self.norm2(hidden_states)
+        norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        norm_hidden_states = norm_hidden_states.unflatten(1, (frames, height, width))
+
+        # Cached conv always supports kv_cache
+        ff_output, kv_cache = self.ff(
+            norm_hidden_states,
+            kv_cache=kv_cache,
+        )
+
+        ff_output = ff_output.flatten(1, 3)
+        hidden_states = hidden_states + gate_mlp * ff_output
+
+        return hidden_states, kv_cache
+
+
+class SanaVideoCausalTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, AttentionMixin):
+    r"""
+    A 3D Transformer model with KV cache support for LongSana autoregressive generation.
+
+    This model extends Sana-Video with causal linear attention and cached convolutions
+    to enable efficient long video generation through chunked processing.
+
+    Args:
+        in_channels (`int`, defaults to `16`):
+            The number of channels in the input.
+        out_channels (`int`, *optional*, defaults to `16`):
+            The number of channels in the output.
+        num_attention_heads (`int`, defaults to `20`):
+            The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, defaults to `112`):
+            The number of channels in each head.
+        num_layers (`int`, defaults to `20`):
+            The number of layers of Transformer blocks to use.
+        num_cross_attention_heads (`int`, *optional*, defaults to `20`):
+            The number of heads to use for cross-attention.
+        cross_attention_head_dim (`int`, *optional*, defaults to `112`):
+            The number of channels in each head for cross-attention.
+        cross_attention_dim (`int`, *optional*, defaults to `2240`):
+            The number of channels in the cross-attention output.
+        caption_channels (`int`, defaults to `2304`):
+            The number of channels in the caption embeddings.
+        mlp_ratio (`float`, defaults to `2.5`):
+            The expansion ratio to use in the GLUMBConv layer.
+        dropout (`float`, defaults to `0.0`):
+            The dropout probability.
+        attention_bias (`bool`, defaults to `False`):
+            Whether to use bias in the attention layer.
+        sample_size (`int`, defaults to `32`):
+            The base size of the input latent.
+        patch_size (`int`, defaults to `1`):
+            The size of the patches to use in the patch embedding layer.
+        norm_elementwise_affine (`bool`, defaults to `False`):
+            Whether to use elementwise affinity in the normalization layer.
+        norm_eps (`float`, defaults to `1e-6`):
+            The epsilon value for the normalization layer.
+        qk_norm (`str`, *optional*, defaults to `None`):
+            The normalization to use for the query and key.
+    """
+
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["SanaVideoCausalTransformerBlock", "SanaModulatedNorm"]
+    _skip_layerwise_casting_patterns = ["patch_embedding", "norm"]
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 16,
+        out_channels: Optional[int] = 16,
+        num_attention_heads: int = 20,
+        attention_head_dim: int = 112,
+        num_layers: int = 20,
+        num_cross_attention_heads: Optional[int] = 20,
+        cross_attention_head_dim: Optional[int] = 112,
+        cross_attention_dim: Optional[int] = 2240,
+        caption_channels: int = 2304,
+        mlp_ratio: float = 2.5,
+        dropout: float = 0.0,
+        attention_bias: bool = False,
+        sample_size: int = 30,
+        patch_size: Tuple[int, int, int] = (1, 2, 2),
+        norm_elementwise_affine: bool = False,
+        norm_eps: float = 1e-6,
+        interpolation_scale: Optional[int] = None,
+        guidance_embeds: bool = False,
+        guidance_embeds_scale: float = 0.1,
+        qk_norm: Optional[str] = "rms_norm_across_heads",
+        rope_max_seq_len: int = 1024,
+    ) -> None:
+        super().__init__()
+
+        out_channels = out_channels or in_channels
+        inner_dim = num_attention_heads * attention_head_dim
+
+        # 1. Patch & position embedding
+        self.rope = WanRotaryPosEmbed(attention_head_dim, patch_size, rope_max_seq_len)
+        self.patch_embedding = nn.Conv3d(in_channels, inner_dim, kernel_size=patch_size, stride=patch_size)
+
+        # 2. Additional condition embeddings
+        if guidance_embeds:
+            self.time_embed = SanaCombinedTimestepGuidanceEmbeddings(inner_dim)
+        else:
+            self.time_embed = AdaLayerNormSingle(inner_dim)
+
+        self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
+        self.caption_norm = RMSNorm(inner_dim, eps=1e-5, elementwise_affine=True)
+
+        # 3. Transformer blocks - use causal versions
+        self.transformer_blocks = nn.ModuleList(
+            [
+                SanaVideoCausalTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    num_cross_attention_heads=num_cross_attention_heads,
+                    cross_attention_head_dim=cross_attention_head_dim,
+                    cross_attention_dim=cross_attention_dim,
+                    attention_bias=attention_bias,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                    mlp_ratio=mlp_ratio,
+                    qk_norm=qk_norm,
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        # 4. Output blocks
+        self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5)
+        self.norm_out = SanaModulatedNorm(inner_dim, elementwise_affine=False, eps=1e-6)
+        self.proj_out = nn.Linear(inner_dim, math.prod(patch_size) * out_channels)
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        timestep: torch.Tensor,
+        guidance: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        kv_caches: Optional[List[SanaBlockKvCache]] = None,
+        return_dict: bool = True,
+    ) -> Union[Tuple[torch.Tensor, ...], SanaVideoCausalTransformer3DModelOutput]:
+        if attention_kwargs is not None:
+            attention_kwargs = attention_kwargs.copy()
+            lora_scale = attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if attention_kwargs is not None and attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
+        if attention_mask is not None and attention_mask.ndim == 2:
+            attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 1. Input
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        p_t, p_h, p_w = self.config.patch_size
+        post_patch_num_frames = num_frames // p_t
+        post_patch_height = height // p_h
+        post_patch_width = width // p_w
+
+        rotary_emb = self.rope(hidden_states)
+
+        hidden_states = self.patch_embedding(hidden_states)
+        hidden_states = hidden_states.flatten(2).transpose(1, 2)
+
+        if guidance is not None:
+            timestep, embedded_timestep = self.time_embed(
+                timestep.flatten(), guidance=guidance, hidden_dtype=hidden_states.dtype
+            )
+        else:
+            timestep, embedded_timestep = self.time_embed(
+                timestep.flatten(), batch_size=batch_size, hidden_dtype=hidden_states.dtype
+            )
+
+        timestep = timestep.view(batch_size, -1, timestep.size(-1))
+        embedded_timestep = embedded_timestep.view(batch_size, -1, embedded_timestep.size(-1))
+
+        encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+        encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+
+        encoder_hidden_states = self.caption_norm(encoder_hidden_states)
+
+        # 2. Transformer blocks with KV cache
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+            # Note: gradient checkpointing doesn't support kv_cache (requires tuple return)
+            if kv_caches is not None:
+                logger.warning("KV cache is not supported with gradient checkpointing. Disabling KV cache.")
+                kv_caches = None
+
+            for index_block, block in enumerate(self.transformer_blocks):
+                hidden_states, _ = self._gradient_checkpointing_func(
+                    block,
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    post_patch_num_frames,
+                    post_patch_height,
+                    post_patch_width,
+                    rotary_emb,
+                    kv_cache=None,
+                )
+        else:
+            for index_block, block in enumerate(self.transformer_blocks):
+                # Get kv_cache for this block if available
+                block_kv_cache = kv_caches[index_block] if kv_caches is not None else None
+
+                hidden_states, block_kv_cache = block(
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    post_patch_num_frames,
+                    post_patch_height,
+                    post_patch_width,
+                    rotary_emb,
+                    kv_cache=block_kv_cache,
+                )
+
+                # Handle return value (could be tensor or tuple)
+                if kv_caches is not None:
+                    kv_caches[index_block] = block_kv_cache
+
+        # 3. Normalization
+        hidden_states = self.norm_out(hidden_states, embedded_timestep, self.scale_shift_table)
+
+        hidden_states = self.proj_out(hidden_states)
+
+        # 5. Unpatchify
+        hidden_states = hidden_states.reshape(
+            batch_size, post_patch_num_frames, post_patch_height, post_patch_width, p_t, p_h, p_w, -1
+        )
+        hidden_states = hidden_states.permute(0, 7, 1, 4, 2, 5, 3, 6)
+        output = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (output, kv_caches)
+
+        return SanaVideoCausalTransformer3DModelOutput(sample=output, kv_cache=kv_caches)

src/diffusers/pipelines/__init__.py

@@ -313,6 +313,7 @@ else:
     ]
     _import_structure["sana_video"] = [
         "SanaVideoPipeline",
+        "LongSanaVideoPipeline",
         "SanaImageToVideoPipeline",
     ]
     _import_structure["semantic_stable_diffusion"] = ["SemanticStableDiffusionPipeline"]
@@ -758,7 +759,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             SanaSprintImg2ImgPipeline,
             SanaSprintPipeline,
         )
-        from .sana_video import SanaImageToVideoPipeline, SanaVideoPipeline
+        from .sana_video import LongSanaVideoPipeline, SanaImageToVideoPipeline, SanaVideoPipeline
         from .semantic_stable_diffusion import SemanticStableDiffusionPipeline
         from .shap_e import ShapEImg2ImgPipeline, ShapEPipeline
         from .stable_audio import StableAudioPipeline, StableAudioProjectionModel

src/diffusers/pipelines/sana/pipeline_sana_sprint.py

@@ -26,7 +26,7 @@ from ...callbacks import MultiPipelineCallbacks, PipelineCallback
 from ...image_processor import PixArtImageProcessor
 from ...loaders import SanaLoraLoaderMixin
 from ...models import AutoencoderDC, SanaTransformer2DModel
-from ...schedulers import DPMSolverMultistepScheduler
+from ...schedulers import SCMScheduler
 from ...utils import (
     BACKENDS_MAPPING,
     USE_PEFT_BACKEND,
@@ -156,7 +156,7 @@ class SanaSprintPipeline(DiffusionPipeline, SanaLoraLoaderMixin):
         text_encoder: Gemma2PreTrainedModel,
         vae: AutoencoderDC,
         transformer: SanaTransformer2DModel,
-        scheduler: DPMSolverMultistepScheduler,
+        scheduler: SCMScheduler,
     ):
         super().__init__()
 

src/diffusers/pipelines/sana_video/__init__.py

@@ -24,6 +24,7 @@ except OptionalDependencyNotAvailable:
 else:
     _import_structure["pipeline_sana_video"] = ["SanaVideoPipeline"]
     _import_structure["pipeline_sana_video_i2v"] = ["SanaImageToVideoPipeline"]
+    _import_structure["pipeline_longsana"] = ["LongSanaVideoPipeline"]
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     try:
@@ -33,6 +34,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     except OptionalDependencyNotAvailable:
         from ...utils.dummy_torch_and_transformers_objects import *
     else:
+        from .pipeline_longsana import LongSanaVideoPipeline
         from .pipeline_sana_video import SanaVideoPipeline
         from .pipeline_sana_video_i2v import SanaImageToVideoPipeline
 else:

src/diffusers/utils/dummy_pt_objects.py

@@ -1383,6 +1383,21 @@ class SanaTransformer2DModel(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class SanaVideoCausalTransformer3DModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class SanaVideoTransformer3DModel(metaclass=DummyObject):
     _backends = ["torch"]
 

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -1727,6 +1727,21 @@ class LEditsPPPipelineStableDiffusionXL(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class LongSanaVideoPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class LTXConditionPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]