update

scripts/convert_mochi_to_diffusers.py

@@ -0,0 +1,187 @@
+import argparse
+from contextlib import nullcontext
+
+import torch
+from accelerate import init_empty_weights
+from safetensors.torch import load_file
+
+# from transformers import T5EncoderModel, T5Tokenizer
+from diffusers import MochiTransformer3DModel
+from diffusers.utils.import_utils import is_accelerate_available
+
+
+CTX = init_empty_weights if is_accelerate_available else nullcontext
+
+TOKENIZER_MAX_LENGTH = 256
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--transformer_checkpoint_path", default=None, type=str)
+# parser.add_argument("--vae_checkpoint_path", default=None, type=str)
+parser.add_argument("--output_path", required=True, type=str)
+parser.add_argument("--push_to_hub", action="store_true", default=False, help="Whether to push to HF Hub after saving")
+parser.add_argument("--text_encoder_cache_dir", type=str, default=None, help="Path to text encoder cache directory")
+parser.add_argument("--dtype", type=str, default=None)
+
+args = parser.parse_args()
+
+
+# This is specific to `AdaLayerNormContinuous`:
+# Diffusers implementation split the linear projection into the scale, shift while Mochi split it into shift, scale
+def swap_scale_shift(weight, dim):
+    shift, scale = weight.chunk(2, dim=0)
+    new_weight = torch.cat([scale, shift], dim=0)
+    return new_weight
+
+
+def convert_mochi_transformer_checkpoint_to_diffusers(ckpt_path):
+    original_state_dict = load_file(ckpt_path, device="cpu")
+    new_state_dict = {}
+
+    # Convert patch_embed
+    new_state_dict["patch_embed.proj.weight"] = original_state_dict.pop("x_embedder.proj.weight")
+    new_state_dict["patch_embed.proj.bias"] = original_state_dict.pop("x_embedder.proj.bias")
+
+    # Convert time_embed
+    new_state_dict["time_embed.timestep_embedder.linear_1.weight"] = original_state_dict.pop("t_embedder.mlp.0.weight")
+    new_state_dict["time_embed.timestep_embedder.linear_1.bias"] = original_state_dict.pop("t_embedder.mlp.0.bias")
+    new_state_dict["time_embed.timestep_embedder.linear_2.weight"] = original_state_dict.pop("t_embedder.mlp.2.weight")
+    new_state_dict["time_embed.timestep_embedder.linear_2.bias"] = original_state_dict.pop("t_embedder.mlp.2.bias")
+    new_state_dict["time_embed.pooler.to_kv.weight"] = original_state_dict.pop("t5_y_embedder.to_kv.weight")
+    new_state_dict["time_embed.pooler.to_kv.bias"] = original_state_dict.pop("t5_y_embedder.to_kv.bias")
+    new_state_dict["time_embed.pooler.to_q.weight"] = original_state_dict.pop("t5_y_embedder.to_q.weight")
+    new_state_dict["time_embed.pooler.to_q.bias"] = original_state_dict.pop("t5_y_embedder.to_q.bias")
+    new_state_dict["time_embed.pooler.to_out.weight"] = original_state_dict.pop("t5_y_embedder.to_out.weight")
+    new_state_dict["time_embed.pooler.to_out.bias"] = original_state_dict.pop("t5_y_embedder.to_out.bias")
+    new_state_dict["time_embed.caption_proj.weight"] = original_state_dict.pop("t5_yproj.weight")
+    new_state_dict["time_embed.caption_proj.bias"] = original_state_dict.pop("t5_yproj.bias")
+
+    # Convert transformer blocks
+    num_layers = 48
+    for i in range(num_layers):
+        block_prefix = f"transformer_blocks.{i}."
+        old_prefix = f"blocks.{i}."
+
+        # norm1
+        new_state_dict[block_prefix + "norm1.linear.weight"] = original_state_dict.pop(old_prefix + "mod_x.weight")
+        new_state_dict[block_prefix + "norm1.linear.bias"] = original_state_dict.pop(old_prefix + "mod_x.bias")
+        if i < num_layers - 1:
+            new_state_dict[block_prefix + "norm1_context.linear.weight"] = original_state_dict.pop(
+                old_prefix + "mod_y.weight"
+            )
+            new_state_dict[block_prefix + "norm1_context.linear.bias"] = original_state_dict.pop(
+                old_prefix + "mod_y.bias"
+            )
+        else:
+            new_state_dict[block_prefix + "norm1_context.linear_1.weight"] = original_state_dict.pop(
+                old_prefix + "mod_y.weight"
+            )
+            new_state_dict[block_prefix + "norm1_context.linear_1.bias"] = original_state_dict.pop(
+                old_prefix + "mod_y.bias"
+            )
+
+        # Visual attention
+        qkv_weight = original_state_dict.pop(old_prefix + "attn.qkv_x.weight")
+        q, k, v = qkv_weight.chunk(3, dim=0)
+
+        new_state_dict[block_prefix + "attn1.to_q.weight"] = q
+        new_state_dict[block_prefix + "attn1.to_k.weight"] = k
+        new_state_dict[block_prefix + "attn1.to_v.weight"] = v
+        new_state_dict[block_prefix + "attn1.norm_q.weight"] = original_state_dict.pop(
+            old_prefix + "attn.q_norm_x.weight"
+        )
+        new_state_dict[block_prefix + "attn1.norm_k.weight"] = original_state_dict.pop(
+            old_prefix + "attn.k_norm_x.weight"
+        )
+        new_state_dict[block_prefix + "attn1.to_out.0.weight"] = original_state_dict.pop(
+            old_prefix + "attn.proj_x.weight"
+        )
+        new_state_dict[block_prefix + "attn1.to_out.0.bias"] = original_state_dict.pop(old_prefix + "attn.proj_x.bias")
+
+        # Context attention
+        qkv_weight = original_state_dict.pop(old_prefix + "attn.qkv_y.weight")
+        q, k, v = qkv_weight.chunk(3, dim=0)
+
+        new_state_dict[block_prefix + "attn1.add_q_proj.weight"] = q
+        new_state_dict[block_prefix + "attn1.add_k_proj.weight"] = k
+        new_state_dict[block_prefix + "attn1.add_v_proj.weight"] = v
+        new_state_dict[block_prefix + "attn1.norm_added_q.weight"] = original_state_dict.pop(
+            old_prefix + "attn.q_norm_y.weight"
+        )
+        new_state_dict[block_prefix + "attn1.norm_added_k.weight"] = original_state_dict.pop(
+            old_prefix + "attn.k_norm_y.weight"
+        )
+        if i < num_layers - 1:
+            new_state_dict[block_prefix + "attn1.to_add_out.weight"] = original_state_dict.pop(
+                old_prefix + "attn.proj_y.weight"
+            )
+            new_state_dict[block_prefix + "attn1.to_add_out.bias"] = original_state_dict.pop(
+                old_prefix + "attn.proj_y.bias"
+            )
+
+        # MLP
+        new_state_dict[block_prefix + "ff.net.0.proj.weight"] = original_state_dict.pop(old_prefix + "mlp_x.w1.weight")
+        new_state_dict[block_prefix + "ff.net.2.weight"] = original_state_dict.pop(old_prefix + "mlp_x.w2.weight")
+        if i < num_layers - 1:
+            new_state_dict[block_prefix + "ff_context.net.0.proj.weight"] = original_state_dict.pop(
+                old_prefix + "mlp_y.w1.weight"
+            )
+            new_state_dict[block_prefix + "ff_context.net.2.weight"] = original_state_dict.pop(
+                old_prefix + "mlp_y.w2.weight"
+            )
+
+    # Output layers
+    new_state_dict["norm_out.linear.weight"] = original_state_dict.pop("final_layer.mod.weight")
+    new_state_dict["norm_out.linear.bias"] = original_state_dict.pop("final_layer.mod.bias")
+    new_state_dict["proj_out.weight"] = original_state_dict.pop("final_layer.linear.weight")
+    new_state_dict["proj_out.bias"] = original_state_dict.pop("final_layer.linear.bias")
+
+    new_state_dict["pos_frequencies"] = original_state_dict.pop("pos_frequencies")
+
+    print("Remaining Keys:", original_state_dict.keys())
+
+    return new_state_dict
+
+
+# def convert_mochi_vae_checkpoint_to_diffusers(ckpt_path, vae_config):
+#     original_state_dict = torch.load(ckpt_path, map_location="cpu")["state_dict"]
+#     return convert_ldm_vae_checkpoint(original_state_dict, vae_config)
+
+
+def main(args):
+    if args.dtype is None:
+        dtype = None
+    if args.dtype == "fp16":
+        dtype = torch.float16
+    elif args.dtype == "bf16":
+        dtype = torch.bfloat16
+    elif args.dtype == "fp32":
+        dtype = torch.float32
+    else:
+        raise ValueError(f"Unsupported dtype: {args.dtype}")
+
+    transformer = None
+    # vae = None
+
+    if args.transformer_checkpoint_path is not None:
+        converted_transformer_state_dict = convert_mochi_transformer_checkpoint_to_diffusers(
+            args.transformer_checkpoint_path
+        )
+        transformer = MochiTransformer3DModel()
+        transformer.load_state_dict(converted_transformer_state_dict, strict=True)
+        if dtype is not None:
+            # Original checkpoint data type will be preserved
+            transformer = transformer.to(dtype=dtype)
+
+    # text_encoder_id = "google/t5-v1_1-xxl"
+    # tokenizer = T5Tokenizer.from_pretrained(text_encoder_id, model_max_length=TOKENIZER_MAX_LENGTH)
+    # text_encoder = T5EncoderModel.from_pretrained(text_encoder_id, cache_dir=args.text_encoder_cache_dir)
+
+    # # Apparently, the conversion does not work anymore without this :shrug:
+    # for param in text_encoder.parameters():
+    #     param.data = param.data.contiguous()
+
+    transformer.save_pretrained("/raid/aryan/mochi-diffusers", subfolder="transformer")
+
+
+if __name__ == "__main__":
+    main(args)

src/diffusers/__init__.py

@@ -100,6 +100,7 @@ else:
             "Kandinsky3UNet",
             "LatteTransformer3DModel",
             "LuminaNextDiT2DModel",
+            "MochiTransformer3DModel",
             "ModelMixin",
             "MotionAdapter",
             "MultiAdapter",
@@ -579,6 +580,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             Kandinsky3UNet,
             LatteTransformer3DModel,
             LuminaNextDiT2DModel,
+            MochiTransformer3DModel,
             ModelMixin,
             MotionAdapter,
             MultiAdapter,

src/diffusers/models/__init__.py

@@ -56,6 +56,7 @@ if is_torch_available():
     _import_structure["transformers.transformer_2d"] = ["Transformer2DModel"]
     _import_structure["transformers.transformer_cogview3plus"] = ["CogView3PlusTransformer2DModel"]
     _import_structure["transformers.transformer_flux"] = ["FluxTransformer2DModel"]
+    _import_structure["transformers.transformer_mochi"] = ["MochiTransformer3DModel"]
     _import_structure["transformers.transformer_sd3"] = ["SD3Transformer2DModel"]
     _import_structure["transformers.transformer_temporal"] = ["TransformerTemporalModel"]
     _import_structure["unets.unet_1d"] = ["UNet1DModel"]
@@ -106,6 +107,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             HunyuanDiT2DModel,
             LatteTransformer3DModel,
             LuminaNextDiT2DModel,
+            MochiTransformer3DModel,
             PixArtTransformer2DModel,
             PriorTransformer,
             SD3Transformer2DModel,

src/diffusers/models/attention_processor.py

@@ -120,6 +120,7 @@ class Attention(nn.Module):
         _from_deprecated_attn_block: bool = False,
         processor: Optional["AttnProcessor"] = None,
         out_dim: int = None,
+        out_context_dim: int = None,
         context_pre_only=None,
         pre_only=False,
         elementwise_affine: bool = True,
@@ -142,6 +143,7 @@ class Attention(nn.Module):
         self.dropout = dropout
         self.fused_projections = False
         self.out_dim = out_dim if out_dim is not None else query_dim
+        self.out_context_dim = out_context_dim if out_context_dim is not None else query_dim
         self.context_pre_only = context_pre_only
         self.pre_only = pre_only
 
@@ -241,7 +243,7 @@ class Attention(nn.Module):
             self.to_out.append(nn.Dropout(dropout))
 
         if self.context_pre_only is not None and not self.context_pre_only:
-            self.to_add_out = nn.Linear(self.inner_dim, self.out_dim, bias=out_bias)
+            self.to_add_out = nn.Linear(self.inner_dim, self.out_context_dim, bias=out_bias)
 
         if qk_norm is not None and added_kv_proj_dim is not None:
             if qk_norm == "fp32_layer_norm":
@@ -1792,6 +1794,7 @@ class FluxAttnProcessor2_0:
             hidden_states = attn.to_out[0](hidden_states)
             # dropout
             hidden_states = attn.to_out[1](hidden_states)
+
             encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
 
             return hidden_states, encoder_hidden_states
@@ -3078,6 +3081,93 @@ class LuminaAttnProcessor2_0:
         return hidden_states
 
 
+class MochiAttnProcessor2_0:
+    """Attention processor used in Mochi."""
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("MochiAttnProcessor2_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: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+
+        query = query.unflatten(2, (attn.heads, -1))
+        key = key.unflatten(2, (attn.heads, -1))
+        value = value.unflatten(2, (attn.heads, -1))
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        encoder_query = attn.add_q_proj(encoder_hidden_states)
+        encoder_key = attn.add_k_proj(encoder_hidden_states)
+        encoder_value = attn.add_v_proj(encoder_hidden_states)
+
+        encoder_query = encoder_query.unflatten(2, (attn.heads, -1))
+        encoder_key = encoder_key.unflatten(2, (attn.heads, -1))
+        encoder_value = encoder_value.unflatten(2, (attn.heads, -1))
+
+        if attn.norm_added_q is not None:
+            encoder_query = attn.norm_added_q(encoder_query)
+        if attn.norm_added_k is not None:
+            encoder_key = attn.norm_added_k(encoder_key)
+
+        if image_rotary_emb is not None:
+
+            def apply_rotary_emb(x, freqs_cos, freqs_sin):
+                x_even = x[..., 0::2].float()
+                x_odd = x[..., 1::2].float()
+
+                cos = (x_even * freqs_cos - x_odd * freqs_sin).to(x.dtype)
+                sin = (x_even * freqs_sin + x_odd * freqs_cos).to(x.dtype)
+
+                return torch.stack([cos, sin], dim=-1).flatten(-2)
+
+            query = apply_rotary_emb(query, *image_rotary_emb)
+            key = apply_rotary_emb(key, *image_rotary_emb)
+
+        query, key, value = query.transpose(1, 2), key.transpose(1, 2), value.transpose(1, 2)
+        encoder_query, encoder_key, encoder_value = (
+            encoder_query.transpose(1, 2),
+            encoder_key.transpose(1, 2),
+            encoder_value.transpose(1, 2),
+        )
+
+        sequence_length = query.size(2)
+        encoder_sequence_length = encoder_query.size(2)
+
+        query = torch.cat([query, encoder_query], dim=2)
+        key = torch.cat([key, encoder_key], dim=2)
+        value = torch.cat([value, encoder_value], dim=2)
+
+        hidden_states = F.scaled_dot_product_attention(query, key, value, dropout_p=0.0, is_causal=False)
+        hidden_states = hidden_states.transpose(1, 2).flatten(2, 3)
+        hidden_states = hidden_states.to(query.dtype)
+
+        hidden_states, encoder_hidden_states = hidden_states.split_with_sizes(
+            (sequence_length, encoder_sequence_length), dim=1
+        )
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
+
+        return hidden_states, encoder_hidden_states
+
+
 class FusedAttnProcessor2_0:
     r"""
     Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). It uses

src/diffusers/models/embeddings.py

@@ -1302,6 +1302,41 @@ class LuminaCombinedTimestepCaptionEmbedding(nn.Module):
         return conditioning
 
 
+class MochiCombinedTimestepCaptionEmbedding(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        pooled_projection_dim: int,
+        text_embed_dim: int,
+        time_embed_dim: int = 256,
+        num_attention_heads: int = 8,
+    ) -> None:
+        super().__init__()
+
+        self.time_proj = Timesteps(num_channels=time_embed_dim, flip_sin_to_cos=True, downscale_freq_shift=0.0)
+        self.timestep_embedder = TimestepEmbedding(in_channels=time_embed_dim, time_embed_dim=embedding_dim)
+        self.pooler = MochiAttentionPool(
+            num_attention_heads=num_attention_heads, embed_dim=text_embed_dim, output_dim=embedding_dim
+        )
+        self.caption_proj = nn.Linear(text_embed_dim, pooled_projection_dim)
+
+    def forward(
+        self,
+        timestep: torch.LongTensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_attention_mask: torch.Tensor,
+        hidden_dtype: Optional[torch.dtype] = None,
+    ):
+        time_proj = self.time_proj(timestep)
+        time_emb = self.timestep_embedder(time_proj.to(dtype=hidden_dtype))
+
+        pooled_projections = self.pooler(encoder_hidden_states, encoder_attention_mask)
+        caption_proj = self.caption_proj(encoder_hidden_states)
+
+        conditioning = time_emb + pooled_projections
+        return conditioning, caption_proj
+
+
 class TextTimeEmbedding(nn.Module):
     def __init__(self, encoder_dim: int, time_embed_dim: int, num_heads: int = 64):
         super().__init__()
@@ -1430,6 +1465,88 @@ class AttentionPooling(nn.Module):
         return a[:, 0, :]  # cls_token
 
 
+class MochiAttentionPool(nn.Module):
+    def __init__(
+        self,
+        num_attention_heads: int,
+        embed_dim: int,
+        output_dim: Optional[int] = None,
+    ) -> None:
+        super().__init__()
+
+        self.output_dim = output_dim or embed_dim
+        self.num_attention_heads = num_attention_heads
+
+        self.to_kv = nn.Linear(embed_dim, 2 * embed_dim)
+        self.to_q = nn.Linear(embed_dim, embed_dim)
+        self.to_out = nn.Linear(embed_dim, self.output_dim)
+
+    @staticmethod
+    def pool_tokens(x: torch.Tensor, mask: torch.Tensor, *, keepdim=False) -> torch.Tensor:
+        """
+        Pool tokens in x using mask.
+
+        NOTE: We assume x does not require gradients.
+
+        Args:
+            x: (B, L, D) tensor of tokens.
+            mask: (B, L) boolean tensor indicating which tokens are not padding.
+
+        Returns:
+            pooled: (B, D) tensor of pooled tokens.
+        """
+        assert x.size(1) == mask.size(1)  # Expected mask to have same length as tokens.
+        assert x.size(0) == mask.size(0)  # Expected mask to have same batch size as tokens.
+        mask = mask[:, :, None].to(dtype=x.dtype)
+        mask = mask / mask.sum(dim=1, keepdim=True).clamp(min=1)
+        pooled = (x * mask).sum(dim=1, keepdim=keepdim)
+        return pooled
+
+    def forward(self, x: torch.Tensor, mask: torch.BoolTensor) -> torch.Tensor:
+        r"""
+        Args:
+            x (`torch.Tensor`):
+                Tensor of shape `(B, S, D)` of input tokens.
+            mask (`torch.Tensor`):
+                Boolean ensor of shape `(B, S)` indicating which tokens are not padding.
+
+        Returns:
+            `torch.Tensor`:
+                `(B, D)` tensor of pooled tokens.
+        """
+        D = x.size(2)
+
+        # Construct attention mask, shape: (B, 1, num_queries=1, num_keys=1+L).
+        attn_mask = mask[:, None, None, :].bool()  # (B, 1, 1, L).
+        attn_mask = F.pad(attn_mask, (1, 0), value=True)  # (B, 1, 1, 1+L).
+
+        # Average non-padding token features. These will be used as the query.
+        x_pool = self.pool_tokens(x, mask, keepdim=True)  # (B, 1, D)
+
+        # Concat pooled features to input sequence.
+        x = torch.cat([x_pool, x], dim=1)  # (B, L+1, D)
+
+        # Compute queries, keys, values. Only the mean token is used to create a query.
+        kv = self.to_kv(x)  # (B, L+1, 2 * D)
+        q = self.to_q(x[:, 0])  # (B, D)
+
+        # Extract heads.
+        head_dim = D // self.num_attention_heads
+        kv = kv.unflatten(2, (2, self.num_attention_heads, head_dim))  # (B, 1+L, 2, H, head_dim)
+        kv = kv.transpose(1, 3)  # (B, H, 2, 1+L, head_dim)
+        k, v = kv.unbind(2)  # (B, H, 1+L, head_dim)
+        q = q.unflatten(1, (self.num_attention_heads, head_dim))  # (B, H, head_dim)
+        q = q.unsqueeze(2)  # (B, H, 1, head_dim)
+
+        # Compute attention.
+        x = F.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask, dropout_p=0.0)  # (B, H, 1, head_dim)
+
+        # Concatenate heads and run output.
+        x = x.squeeze(2).flatten(1, 2)  # (B, D = H * head_dim)
+        x = self.to_out(x)
+        return x
+
+
 def get_fourier_embeds_from_boundingbox(embed_dim, box):
     """
     Args:

src/diffusers/models/normalization.py

@@ -237,6 +237,33 @@ class LuminaRMSNormZero(nn.Module):
         return x, gate_msa, scale_mlp, gate_mlp
 
 
+class MochiRMSNormZero(nn.Module):
+    r"""
+    Adaptive RMS Norm used in Mochi.
+
+    Parameters:
+        embedding_dim (`int`): The size of each embedding vector.
+    """
+
+    def __init__(
+        self, embedding_dim: int, hidden_dim: int, eps: float = 1e-5, elementwise_affine: bool = False
+    ) -> None:
+        super().__init__()
+
+        self.silu = nn.SiLU()
+        self.linear = nn.Linear(embedding_dim, hidden_dim)
+        self.norm = RMSNorm(embedding_dim, eps=eps, elementwise_affine=elementwise_affine)
+
+    def forward(
+        self, hidden_states: torch.Tensor, emb: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        emb = self.linear(self.silu(emb))
+        scale_msa, gate_msa, scale_mlp, gate_mlp = emb.chunk(4, dim=1)
+        hidden_states = self.norm(hidden_states) * (1 + scale_msa[:, None])
+
+        return hidden_states, gate_msa, scale_mlp, gate_mlp
+
+
 class AdaLayerNormSingle(nn.Module):
     r"""
     Norm layer adaptive layer norm single (adaLN-single).
@@ -358,20 +385,21 @@ class LuminaLayerNormContinuous(nn.Module):
         out_dim: Optional[int] = None,
     ):
         super().__init__()
+
         # AdaLN
         self.silu = nn.SiLU()
         self.linear_1 = nn.Linear(conditioning_embedding_dim, embedding_dim, bias=bias)
+
         if norm_type == "layer_norm":
             self.norm = LayerNorm(embedding_dim, eps, elementwise_affine, bias)
+        if norm_type == "rms_norm":
+            self.norm = RMSNorm(embedding_dim, eps=eps, elementwise_affine=elementwise_affine)
         else:
             raise ValueError(f"unknown norm_type {norm_type}")
-        # linear_2
+
+        self.linear_2 = None
         if out_dim is not None:
-            self.linear_2 = nn.Linear(
-                embedding_dim,
-                out_dim,
-                bias=bias,
-            )
+            self.linear_2 = nn.Linear(embedding_dim, out_dim, bias=bias)
 
     def forward(
         self,

src/diffusers/models/transformers/__init__.py

@@ -16,5 +16,6 @@ if is_torch_available():
     from .transformer_2d import Transformer2DModel
     from .transformer_cogview3plus import CogView3PlusTransformer2DModel
     from .transformer_flux import FluxTransformer2DModel
+    from .transformer_mochi import MochiTransformer3DModel
     from .transformer_sd3 import SD3Transformer2DModel
     from .transformer_temporal import TransformerTemporalModel

src/diffusers/models/transformers/transformer_mochi.py

@@ -0,0 +1,306 @@
+# Copyright 2024 The Genmo team and The HuggingFace 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.
+
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils import logging
+from ...utils.torch_utils import maybe_allow_in_graph
+from ..attention import FeedForward
+from ..attention_processor import Attention, MochiAttnProcessor2_0
+from ..embeddings import MochiCombinedTimestepCaptionEmbedding, PatchEmbed
+from ..modeling_outputs import Transformer2DModelOutput
+from ..modeling_utils import ModelMixin
+from ..normalization import AdaLayerNormContinuous, LuminaLayerNormContinuous, MochiRMSNormZero, RMSNorm
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@maybe_allow_in_graph
+class MochiTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        pooled_projection_dim: int,
+        qk_norm: str = "rms_norm",
+        activation_fn: str = "swiglu",
+        context_pre_only: bool = True,
+        eps: float = 1e-6,
+    ) -> None:
+        super().__init__()
+
+        self.context_pre_only = context_pre_only
+        self.ff_inner_dim = (4 * dim * 2) // 3
+        self.ff_context_inner_dim = (4 * pooled_projection_dim * 2) // 3
+
+        self.norm1 = MochiRMSNormZero(dim, 4 * dim, eps=eps, elementwise_affine=False)
+
+        if not context_pre_only:
+            self.norm1_context = MochiRMSNormZero(dim, 4 * pooled_projection_dim, eps=eps, elementwise_affine=False)
+        else:
+            self.norm1_context = LuminaLayerNormContinuous(
+                embedding_dim=pooled_projection_dim,
+                conditioning_embedding_dim=dim,
+                eps=eps,
+                elementwise_affine=False,
+                norm_type="rms_norm",
+                out_dim=None,
+            )
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            cross_attention_dim=None,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            bias=False,
+            qk_norm=qk_norm,
+            added_kv_proj_dim=pooled_projection_dim,
+            added_proj_bias=False,
+            out_dim=dim,
+            out_context_dim=pooled_projection_dim,
+            context_pre_only=context_pre_only,
+            processor=MochiAttnProcessor2_0(),
+            eps=eps,
+            elementwise_affine=True,
+        )
+
+        self.norm2 = RMSNorm(dim, eps=eps, elementwise_affine=False)
+        self.norm2_context = RMSNorm(pooled_projection_dim, eps=eps, elementwise_affine=False)
+
+        self.norm3 = RMSNorm(dim, eps=eps, elementwise_affine=False)
+        self.norm3_context = RMSNorm(pooled_projection_dim, eps=eps, elementwise_affine=False)
+
+        self.ff = FeedForward(dim, inner_dim=self.ff_inner_dim, activation_fn=activation_fn, bias=False)
+        self.ff_context = None
+        if not context_pre_only:
+            self.ff_context = FeedForward(
+                pooled_projection_dim, inner_dim=self.ff_context_inner_dim, activation_fn=activation_fn, bias=False
+            )
+
+        self.norm4 = RMSNorm(dim, eps=eps, elementwise_affine=False)
+        self.norm4_context = RMSNorm(pooled_projection_dim, eps=eps, elementwise_affine=False)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        breakpoint()
+        norm_hidden_states, gate_msa, scale_mlp, gate_mlp = self.norm1(hidden_states, temb)
+
+        if not self.context_pre_only:
+            norm_encoder_hidden_states, enc_gate_msa, enc_scale_mlp, enc_gate_mlp = self.norm1_context(
+                encoder_hidden_states, temb
+            )
+        else:
+            norm_encoder_hidden_states = self.norm1_context(encoder_hidden_states, temb)
+
+        attn_hidden_states, context_attn_hidden_states = self.attn1(
+            hidden_states=norm_hidden_states,
+            encoder_hidden_states=norm_encoder_hidden_states,
+            image_rotary_emb=image_rotary_emb,
+        )
+
+        hidden_states = hidden_states + self.norm2(attn_hidden_states) * torch.tanh(gate_msa).unsqueeze(1)
+        norm_hidden_states = self.norm3(hidden_states) * (1 + scale_mlp.unsqueeze(1))
+
+        if not self.context_pre_only:
+            encoder_hidden_states = encoder_hidden_states + self.norm2_context(
+                context_attn_hidden_states
+            ) * torch.tanh(enc_gate_msa).unsqueeze(1)
+            norm_encoder_hidden_states = encoder_hidden_states + self.norm3_context(encoder_hidden_states) * (
+                1 + enc_scale_mlp.unsqueeze(1)
+            )
+
+        ff_output = self.ff(norm_hidden_states)
+        hidden_states = hidden_states + ff_output * torch.tanh(gate_mlp).unsqueeze(1)
+
+        if not self.context_pre_only:
+            context_ff_output = self.ff_context(norm_encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states + context_ff_output * torch.tanh(enc_gate_mlp).unsqueeze(0)
+
+        return hidden_states, encoder_hidden_states
+
+
+class MochiRoPE(nn.Module):
+    def __init__(self, base_height: int = 192, base_width: int = 192, theta: float = 10000.0) -> None:
+        super().__init__()
+
+        self.target_area = base_height * base_width
+
+    def _centers(self, start, stop, num, device, dtype) -> torch.Tensor:
+        edges = torch.linspace(start, stop, num + 1, device=device, dtype=dtype)
+        return (edges[:-1] + edges[1:]) / 2
+
+    def _get_positions(
+        self,
+        num_frames: int,
+        height: int,
+        width: int,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ) -> torch.Tensor:
+        scale = (self.target_area / (height * width)) ** 0.5
+
+        t = torch.arange(num_frames, device=device, dtype=dtype)
+        h = self._centers(-height * scale / 2, height * scale / 2, height, device, dtype)
+        w = self._centers(-width * scale / 2, width * scale / 2, width, device, dtype)
+
+        grid_t, grid_h, grid_w = torch.meshgrid(t, h, w, indexing="ij")
+
+        positions = torch.stack([grid_t, grid_h, grid_w], dim=-1).view(-1, 3)
+        return positions
+
+    def _create_rope(self, freqs: torch.Tensor, pos: torch.Tensor) -> torch.Tensor:
+        freqs = torch.einsum("nd,dhf->nhf", pos, freqs)
+        freqs_cos = torch.cos(freqs)
+        freqs_sin = torch.sin(freqs)
+        return freqs_cos, freqs_sin
+
+    def forward(
+        self,
+        pos_frequencies: torch.Tensor,
+        num_frames: int,
+        height: int,
+        width: int,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        pos = self._get_positions(num_frames, height, width, device, dtype)
+        rope_cos, rope_sin = self._create_rope(pos_frequencies, pos)
+        return rope_cos, rope_sin
+
+
+@maybe_allow_in_graph
+class MochiTransformer3DModel(ModelMixin, ConfigMixin):
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        patch_size: int = 2,
+        num_attention_heads: int = 24,
+        attention_head_dim: int = 128,
+        num_layers: int = 48,
+        pooled_projection_dim: int = 1536,
+        in_channels: int = 12,
+        out_channels: Optional[int] = None,
+        qk_norm: str = "rms_norm",
+        text_embed_dim: int = 4096,
+        time_embed_dim: int = 256,
+        activation_fn: str = "swiglu",
+        max_sequence_length: int = 256,
+    ) -> None:
+        super().__init__()
+
+        inner_dim = num_attention_heads * attention_head_dim
+        out_channels = out_channels or in_channels
+
+        self.patch_embed = PatchEmbed(
+            patch_size=patch_size,
+            in_channels=in_channels,
+            embed_dim=inner_dim,
+            pos_embed_type=None,
+        )
+
+        self.time_embed = MochiCombinedTimestepCaptionEmbedding(
+            embedding_dim=inner_dim,
+            pooled_projection_dim=pooled_projection_dim,
+            text_embed_dim=text_embed_dim,
+            time_embed_dim=time_embed_dim,
+            num_attention_heads=8,
+        )
+
+        self.pos_frequencies = nn.Parameter(torch.empty(3, num_attention_heads, attention_head_dim // 2))
+        self.rope = MochiRoPE()
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                MochiTransformerBlock(
+                    dim=inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    pooled_projection_dim=pooled_projection_dim,
+                    qk_norm=qk_norm,
+                    activation_fn=activation_fn,
+                    context_pre_only=i == num_layers - 1,
+                )
+                for i in range(num_layers)
+            ]
+        )
+
+        self.norm_out = AdaLayerNormContinuous(
+            inner_dim, inner_dim, elementwise_affine=False, eps=1e-6, norm_type="layer_norm"
+        )
+        self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * out_channels)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        timestep: torch.LongTensor,
+        encoder_attention_mask: torch.Tensor,
+        return_dict: bool = True,
+    ) -> torch.Tensor:
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        p = self.config.patch_size
+
+        post_patch_height = height // p
+        post_patch_width = width // p
+
+        temb, encoder_hidden_states = self.time_embed(
+            timestep, encoder_hidden_states, encoder_attention_mask, hidden_dtype=hidden_states.dtype
+        )
+
+        hidden_states = hidden_states.permute(0, 2, 1, 3, 4).flatten(0, 1)
+        hidden_states = self.patch_embed(hidden_states)
+        hidden_states = hidden_states.unflatten(0, (batch_size, -1)).flatten(1, 2)
+
+        image_rotary_emb = self.rope(
+            self.pos_frequencies,
+            num_frames,
+            post_patch_height,
+            post_patch_width,
+            device=hidden_states.device,
+            dtype=torch.float32,
+        )
+
+        for i, block in enumerate(self.transformer_blocks):
+            hidden_states, encoder_hidden_states = block(
+                hidden_states=hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                temb=temb,
+                image_rotary_emb=image_rotary_emb,
+            )
+
+        # TODO(aryan): do something with self.pos_frequencies
+        hidden_states = self.norm_out(hidden_states, temb)
+        hidden_states = self.proj_out(hidden_states)
+
+        hidden_states = hidden_states.reshape(batch_size, num_frames, post_patch_height, post_patch_width, p, p, -1)
+        hidden_states = hidden_states.permute(0, 6, 1, 2, 4, 3, 5)
+        output = hidden_states.reshape(batch_size, -1, num_frames, height, width)
+
+        if not return_dict:
+            return (output,)
+        return Transformer2DModelOutput(sample=output)

src/diffusers/pipelines/mochi/pipeline_mochi.py

@@ -0,0 +1,669 @@
+# Copyright 2024 Black Forest Labs and The HuggingFace 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 inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import torch
+from transformers import T5EncoderModel, T5TokenizerFast
+
+from ...loaders import TextualInversionLoaderMixin
+from ...models.autoencoders import AutoencoderKL
+from ...models.transformers import MochiTransformer3DModel
+from ...schedulers import FlowMatchEulerDiscreteScheduler
+from ...utils import (
+    is_torch_xla_available,
+    logging,
+    replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ...video_processor import VideoProcessor
+from ..pipeline_utils import DiffusionPipeline
+from .pipeline_output import MochiPipelineOutput
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import MochiPipeline
+
+        >>> pipe = MochiPipeline.from_pretrained("black-forest-labs/mochi.1-schnell", torch_dtype=torch.bfloat16)
+        >>> pipe.to("cuda")
+        >>> prompt = "A cat holding a sign that says hello world"
+        >>> # Depending on the variant being used, the pipeline call will slightly vary.
+        >>> # Refer to the pipeline documentation for more details.
+        >>> image = pipe(prompt, num_inference_steps=4, guidance_scale=0.0).images[0]
+        >>> image.save("mochi.png")
+        ```
+"""
+
+
+def calculate_shift(
+    image_seq_len,
+    base_seq_len: int = 256,
+    max_seq_len: int = 4096,
+    base_shift: float = 0.5,
+    max_shift: float = 1.16,
+):
+    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
+    b = base_shift - m * base_seq_len
+    mu = image_seq_len * m + b
+    return mu
+
+
+# from: https://github.com/genmoai/models/blob/075b6e36db58f1242921deff83a1066887b9c9e1/src/mochi_preview/infer.py#L77
+def linear_quadratic_schedule(num_steps, threshold_noise, linear_steps=None):
+    if linear_steps is None:
+        linear_steps = num_steps // 2
+    linear_sigma_schedule = [i * threshold_noise / linear_steps for i in range(linear_steps)]
+    threshold_noise_step_diff = linear_steps - threshold_noise * num_steps
+    quadratic_steps = num_steps - linear_steps
+    quadratic_coef = threshold_noise_step_diff / (linear_steps * quadratic_steps**2)
+    linear_coef = threshold_noise / linear_steps - 2 * threshold_noise_step_diff / (quadratic_steps**2)
+    const = quadratic_coef * (linear_steps**2)
+    quadratic_sigma_schedule = [
+        quadratic_coef * (i**2) + linear_coef * i + const for i in range(linear_steps, num_steps)
+    ]
+    sigma_schedule = linear_sigma_schedule + quadratic_sigma_schedule + [1.0]
+    sigma_schedule = [1.0 - x for x in sigma_schedule]
+    return sigma_schedule
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class MochiPipeline(DiffusionPipeline, TextualInversionLoaderMixin):
+    r"""
+    The mochi pipeline for text-to-image generation.
+
+    Reference: https://blackforestlabs.ai/announcing-black-forest-labs/
+
+    Args:
+        transformer ([`mochiTransformer2DModel`]):
+            Conditional Transformer (MMDiT) architecture to denoise the encoded image latents.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+        text_encoder ([`CLIPTextModel`]):
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        text_encoder ([`T5EncoderModel`]):
+            [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
+            the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
+        tokenizer (`T5TokenizerFast`):
+            Second Tokenizer of class
+            [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast).
+    """
+
+    model_cpu_offload_seq = "text_encoder->transformer->vae"
+    _optional_components = []
+    _callback_tensor_inputs = ["latents", "prompt_embeds"]
+
+    def __init__(
+        self,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        vae: AutoencoderKL,
+        text_encoder: T5EncoderModel,
+        tokenizer: T5TokenizerFast,
+        transformer: MochiTransformer3DModel,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+        )
+        # TODO: determine these scaling factors from model parameters
+        self.vae_spatial_scale_factor = 8
+        self.vae_temporal_scale_factor = 6
+        self.patch_size = 2
+
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor)
+        self.tokenizer_max_length = (
+            self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77
+        )
+        self.default_height = 480
+        self.default_width = 848
+
+    def _get_t5_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]] = None,
+        num_videos_per_prompt: int = 1,
+        max_sequence_length: int = 512,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        batch_size = len(prompt)
+
+        if isinstance(self, TextualInversionLoaderMixin):
+            prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            return_length=False,
+            return_overflowing_tokens=False,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids
+        prompt_attention_mask = text_inputs.attention_mask
+        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1])
+            logger.warning(
+                "The following part of your input was truncated because `max_sequence_length` is set to "
+                f" {max_sequence_length} tokens: {removed_text}"
+            )
+
+        prompt_embeds = self.text_encoder(
+            text_input_ids.to(device), attention_mask=prompt_attention_mask.to(device), output_hidden_states=False
+        ).last_hidden_state
+
+        dtype = self.text_encoder.dtype
+        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+        _, seq_len, _ = prompt_embeds.shape
+
+        # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+        return prompt_embeds
+
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        device: Optional[torch.device] = None,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        max_sequence_length: int = 512,
+        do_classifier_free_guidance: bool = True,
+    ):
+        r"""
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            device: (`torch.device`):
+                torch device
+            num_videos_per_prompt (`int`):
+                number of images that should be generated per prompt
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            lora_scale (`float`, *optional*):
+                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+        """
+        device = device or self._execution_device
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+            )
+
+        prompt_embeds = prompt_embeds.to(self.text_encoder.dtype)
+
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            # normalize str to list
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+
+            negative_prompt_embeds = self._get_t5_prompt_embeds(
+                prompt=negative_prompt,
+                num_videos_per_prompt=num_videos_per_prompt,
+                max_sequence_length=max_sequence_length,
+                device=device,
+            )
+            negative_prompt_embeds = negative_prompt_embeds.to(self.text_encoder.dtype)
+
+        return prompt_embeds, negative_prompt_embeds
+
+    def check_inputs(
+        self,
+        prompt,
+        height,
+        width,
+        prompt_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+        max_sequence_length=None,
+    ):
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+        if max_sequence_length is not None and max_sequence_length > 512:
+            raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}")
+
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    def prepare_latents(
+        self,
+        batch_size,
+        num_channels_latents,
+        height,
+        width,
+        num_frames,
+        dtype,
+        device,
+        generator,
+        latents=None,
+    ):
+        height = height // self.vae_spatial_scale_factor
+        width = width // self.vae_spatial_scale_factor
+        num_frames = (num_frames - 1) // (self.vae_temporal_scale_factor + 1)
+
+        shape = (batch_size, num_channels_latents, num_frames, height, width)
+
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        return latents
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1.0
+
+    @property
+    def joint_attention_kwargs(self):
+        return self._joint_attention_kwargs
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_frames: int = 16,
+        num_inference_steps: int = 28,
+        timesteps: List[int] = None,
+        guidance_scale: float = 3.5,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 512,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image. This is set to 1024 by default for the best results.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image. This is set to 1024 by default for the best results.
+            num_frames (`int`, defaults to 16):
+                The number of video frames to generate
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            guidance_scale (`float`, *optional*, defaults to 7.0):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.mochi.MochiPipelineOutput`] instead of a plain tuple.
+            joint_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.mochi.MochiPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if
+            `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the
+            generated images.
+        """
+        height = height or self.default_height
+        width = width or self.default_width
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            prompt_embeds=prompt_embeds,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+            max_sequence_length=max_sequence_length,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._joint_attention_kwargs = joint_attention_kwargs
+        self._interrupt = False
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        lora_scale = (
+            self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None
+        )
+        (prompt_embeds, negative_prompt_embeds) = self.encode_prompt(
+            prompt=prompt,
+            prompt_embeds=prompt_embeds,
+            device=device,
+            num_videos_per_prompt=num_videos_per_prompt,
+            max_sequence_length=max_sequence_length,
+            lora_scale=lora_scale,
+        )
+
+        # 4. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_frames,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 5. Prepare timestep
+
+        # from https://github.com/genmoai/models/blob/075b6e36db58f1242921deff83a1066887b9c9e1/src/mochi_preview/infer.py#L77
+        threshold_noise = 0.025
+        sigmas = linear_quadratic_schedule(num_inference_steps, threshold_noise)
+
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            timesteps,
+            sigmas,
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
+
+        if self.do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+
+        # 6. Denoising loop
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latents.shape[0]).to(latents.dtype)
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    timestep=timestep,
+                    encoder_hidden_states=prompt_embeds,
+                    joint_attention_kwargs=self.joint_attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                if self.do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents_dtype = latents.dtype
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if latents.dtype != latents_dtype:
+                    if torch.backends.mps.is_available():
+                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
+                        latents = latents.to(latents_dtype)
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        if output_type == "latent":
+            video = latents
+
+        else:
+            video = self.vae.decode(latents, return_dict=False)[0]
+            video = self.video_processor.postprocess(video, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return MochiPipelineOutput(frames=video)

src/diffusers/pipelines/mochi/pipeline_output.py

@@ -0,0 +1,20 @@
+from dataclasses import dataclass
+
+import torch
+
+from diffusers.utils import BaseOutput
+
+
+@dataclass
+class MochiPipelineOutput(BaseOutput):
+    r"""
+    Output class for CogVideo pipelines.
+
+    Args:
+        frames (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]):
+            List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing
+            denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape
+            `(batch_size, num_frames, channels, height, width)`.
+    """
+
+    frames: torch.Tensor

src/diffusers/utils/dummy_pt_objects.py

@@ -347,6 +347,21 @@ class LuminaNextDiT2DModel(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class MochiTransformer3DModel(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 ModelMixin(metaclass=DummyObject):
     _backends = ["torch"]