Cogvideox-5B Model adapter change (#9203)

* draft of embedding

---------

Co-authored-by: Aryan <aryan@huggingface.co>

Author: 2 people

docs/source/en/api/pipelines/cogvideox.md

@@ -29,6 +29,10 @@ Make sure to check out the Schedulers [guide](../../using-diffusers/schedulers.m
 
 This pipeline was contributed by [zRzRzRzRzRzRzR](https://github.com/zRzRzRzRzRzRzR). The original codebase can be found [here](https://huggingface.co/THUDM). The original weights can be found under [hf.co/THUDM](https://huggingface.co/THUDM).
 
+There are two models available that can be used with the CogVideoX pipeline:
+- [`THUDM/CogVideoX-2b`](https://huggingface.co/THUDM/CogVideoX-2b)
+- [`THUDM/CogVideoX-5b`](https://huggingface.co/THUDM/CogVideoX-5b)
+
 ## Inference
 
 Use [`torch.compile`](https://huggingface.co/docs/diffusers/main/en/tutorials/fast_diffusion#torchcompile) to reduce the inference latency.
@@ -68,7 +72,7 @@ With torch.compile(): Average inference time: 76.27 seconds.
 
 ### Memory optimization
 
-CogVideoX requires about 19 GB of GPU memory to decode 49 frames (6 seconds of video at 8 FPS) with output resolution 720x480 (W x H), which makes it not possible to run on consumer GPUs or free-tier T4 Colab. The following memory optimizations could be used to reduce the memory footprint. For replication, you can refer to [this](https://gist.github.com/a-r-r-o-w/3959a03f15be5c9bd1fe545b09dfcc93) script.
+CogVideoX-2b requires about 19 GB of GPU memory to decode 49 frames (6 seconds of video at 8 FPS) with output resolution 720x480 (W x H), which makes it not possible to run on consumer GPUs or free-tier T4 Colab. The following memory optimizations could be used to reduce the memory footprint. For replication, you can refer to [this](https://gist.github.com/a-r-r-o-w/3959a03f15be5c9bd1fe545b09dfcc93) script.
 
 - `pipe.enable_model_cpu_offload()`:
   - Without enabling cpu offloading, memory usage is `33 GB`

scripts/convert_cogvideox_to_diffusers.py

@@ -86,6 +86,9 @@ def replace_up_keys_inplace(key: str, state_dict: Dict[str, Any]):
     "key_layernorm_list": reassign_query_key_layernorm_inplace,
     "adaln_layer.adaLN_modulations": reassign_adaln_norm_inplace,
     "embed_tokens": remove_keys_inplace,
+    "freqs_sin": remove_keys_inplace,
+    "freqs_cos": remove_keys_inplace,
+    "position_embedding": remove_keys_inplace,
 }
 
 VAE_KEYS_RENAME_DICT = {
@@ -123,11 +126,21 @@ def update_state_dict_inplace(state_dict: Dict[str, Any], old_key: str, new_key:
     state_dict[new_key] = state_dict.pop(old_key)
 
 
-def convert_transformer(ckpt_path: str):
+def convert_transformer(
+    ckpt_path: str,
+    num_layers: int,
+    num_attention_heads: int,
+    use_rotary_positional_embeddings: bool,
+    dtype: torch.dtype,
+):
     PREFIX_KEY = "model.diffusion_model."
 
     original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", mmap=True))
-    transformer = CogVideoXTransformer3DModel()
+    transformer = CogVideoXTransformer3DModel(
+        num_layers=num_layers,
+        num_attention_heads=num_attention_heads,
+        use_rotary_positional_embeddings=use_rotary_positional_embeddings,
+    ).to(dtype=dtype)
 
     for key in list(original_state_dict.keys()):
         new_key = key[len(PREFIX_KEY) :]
@@ -145,9 +158,9 @@ def convert_transformer(ckpt_path: str):
     return transformer
 
 
-def convert_vae(ckpt_path: str):
+def convert_vae(ckpt_path: str, scaling_factor: float, dtype: torch.dtype):
     original_state_dict = get_state_dict(torch.load(ckpt_path, map_location="cpu", mmap=True))
-    vae = AutoencoderKLCogVideoX()
+    vae = AutoencoderKLCogVideoX(scaling_factor=scaling_factor).to(dtype=dtype)
 
     for key in list(original_state_dict.keys()):
         new_key = key[:]
@@ -172,13 +185,26 @@ def get_args():
     )
     parser.add_argument("--vae_ckpt_path", type=str, default=None, help="Path to original vae checkpoint")
     parser.add_argument("--output_path", type=str, required=True, help="Path where converted model should be saved")
-    parser.add_argument("--fp16", action="store_true", default=True, help="Whether to save the model weights in fp16")
+    parser.add_argument("--fp16", action="store_true", default=False, help="Whether to save the model weights in fp16")
+    parser.add_argument("--bf16", action="store_true", default=False, help="Whether to save the model weights in bf16")
     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"
     )
+    # For CogVideoX-2B, num_layers is 30. For 5B, it is 42
+    parser.add_argument("--num_layers", type=int, default=30, help="Number of transformer blocks")
+    # For CogVideoX-2B, num_attention_heads is 30. For 5B, it is 48
+    parser.add_argument("--num_attention_heads", type=int, default=30, help="Number of attention heads")
+    # For CogVideoX-2B, use_rotary_positional_embeddings is False. For 5B, it is True
+    parser.add_argument(
+        "--use_rotary_positional_embeddings", action="store_true", default=False, help="Whether to use RoPE or not"
+    )
+    # For CogVideoX-2B, scaling_factor is 1.15258426. For 5B, it is 0.7
+    parser.add_argument("--scaling_factor", type=float, default=1.15258426, help="Scaling factor in the VAE")
+    # For CogVideoX-2B, snr_shift_scale is 3.0. For 5B, it is 1.0
+    parser.add_argument("--snr_shift_scale", type=float, default=3.0, help="Scaling factor in the VAE")
     return parser.parse_args()
 
 
@@ -188,18 +214,33 @@ def get_args():
     transformer = None
     vae = None
 
+    if args.fp16 and args.bf16:
+        raise ValueError("You cannot pass both --fp16 and --bf16 at the same time.")
+
+    dtype = torch.float16 if args.fp16 else torch.bfloat16 if args.bf16 else torch.float32
+
     if args.transformer_ckpt_path is not None:
-        transformer = convert_transformer(args.transformer_ckpt_path)
+        transformer = convert_transformer(
+            args.transformer_ckpt_path,
+            args.num_layers,
+            args.num_attention_heads,
+            args.use_rotary_positional_embeddings,
+            dtype,
+        )
     if args.vae_ckpt_path is not None:
-        vae = convert_vae(args.vae_ckpt_path)
+        vae = convert_vae(args.vae_ckpt_path, args.scaling_factor, 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 any more without this :shrug:
+    for param in text_encoder.parameters():
+        param.data = param.data.contiguous()
+
     scheduler = CogVideoXDDIMScheduler.from_config(
         {
-            "snr_shift_scale": 3.0,
+            "snr_shift_scale": args.snr_shift_scale,
             "beta_end": 0.012,
             "beta_schedule": "scaled_linear",
             "beta_start": 0.00085,
@@ -208,7 +249,7 @@ def get_args():
             "prediction_type": "v_prediction",
             "rescale_betas_zero_snr": True,
             "set_alpha_to_one": True,
-            "timestep_spacing": "linspace",
+            "timestep_spacing": "trailing",
         }
     )
 
@@ -218,5 +259,10 @@ def get_args():
 
     if args.fp16:
         pipe = pipe.to(dtype=torch.float16)
+    if args.bf16:
+        pipe = pipe.to(dtype=torch.bfloat16)
 
+    # We don't use variant here because the model must be run in fp16 (2B) or bf16 (5B). It would be weird
+    # for users to specify variant when the default is not fp32 and they want to run with the correct default (which
+    # is either fp16/bf16 here).
     pipe.save_pretrained(args.output_path, safe_serialization=True, push_to_hub=args.push_to_hub)

src/diffusers/models/attention_processor.py

@@ -1783,6 +1783,148 @@ def __call__(
             return hidden_states
 
 
+class CogVideoXAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention for the CogVideoX model. It applies a rotary embedding on
+    query and key vectors, but does not include spatial normalization.
+    """
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("CogVideoXAttnProcessor 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:
+        text_seq_length = encoder_hidden_states.size(1)
+
+        hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+
+        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)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # Apply RoPE if needed
+        if image_rotary_emb is not None:
+            from .embeddings import apply_rotary_emb
+
+            query[:, :, text_seq_length:] = apply_rotary_emb(query[:, :, text_seq_length:], image_rotary_emb)
+            if not attn.is_cross_attention:
+                key[:, :, text_seq_length:] = apply_rotary_emb(key[:, :, text_seq_length:], image_rotary_emb)
+
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        encoder_hidden_states, hidden_states = hidden_states.split(
+            [text_seq_length, hidden_states.size(1) - text_seq_length], dim=1
+        )
+        return hidden_states, encoder_hidden_states
+
+
+class FusedCogVideoXAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention for the CogVideoX model. It applies a rotary embedding on
+    query and key vectors, but does not include spatial normalization.
+    """
+
+    def __init__(self):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("CogVideoXAttnProcessor 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:
+        text_seq_length = encoder_hidden_states.size(1)
+
+        hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
+
+        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)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        qkv = attn.to_qkv(hidden_states)
+        split_size = qkv.shape[-1] // 3
+        query, key, value = torch.split(qkv, split_size, dim=-1)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # Apply RoPE if needed
+        if image_rotary_emb is not None:
+            from .embeddings import apply_rotary_emb
+
+            query[:, :, text_seq_length:] = apply_rotary_emb(query[:, :, text_seq_length:], image_rotary_emb)
+            if not attn.is_cross_attention:
+                key[:, :, text_seq_length:] = apply_rotary_emb(key[:, :, text_seq_length:], image_rotary_emb)
+
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        encoder_hidden_states, hidden_states = hidden_states.split(
+            [text_seq_length, hidden_states.size(1) - text_seq_length], dim=1
+        )
+        return hidden_states, encoder_hidden_states
+
+
 class XFormersAttnAddedKVProcessor:
     r"""
     Processor for implementing memory efficient attention using xFormers.

src/diffusers/models/autoencoders/autoencoder_kl_cogvideox.py

@@ -902,7 +902,7 @@ class AutoencoderKLCogVideoX(ModelMixin, ConfigMixin, FromOriginalModelMixin):
             Tuple of block output channels.
         act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
         sample_size (`int`, *optional*, defaults to `32`): Sample input size.
-        scaling_factor (`float`, *optional*, defaults to 0.18215):
+        scaling_factor (`float`, *optional*, defaults to `1.15258426`):
             The component-wise standard deviation of the trained latent space computed using the first batch of the
             training set. This is used to scale the latent space to have unit variance when training the diffusion
             model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the

src/diffusers/models/embeddings.py

@@ -374,6 +374,90 @@ def forward(self, text_embeds: torch.Tensor, image_embeds: torch.Tensor):
         return embeds
 
 
+def get_3d_rotary_pos_embed(
+    embed_dim, crops_coords, grid_size, temporal_size, theta: int = 10000, use_real: bool = True
+) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+    """
+    RoPE for video tokens with 3D structure.
+
+    Args:
+    embed_dim: (`int`):
+        The embedding dimension size, corresponding to hidden_size_head.
+    crops_coords (`Tuple[int]`):
+        The top-left and bottom-right coordinates of the crop.
+    grid_size (`Tuple[int]`):
+        The grid size of the spatial positional embedding (height, width).
+    temporal_size (`int`):
+        The size of the temporal dimension.
+    theta (`float`):
+        Scaling factor for frequency computation.
+    use_real (`bool`):
+        If True, return real part and imaginary part separately. Otherwise, return complex numbers.
+
+    Returns:
+        `torch.Tensor`: positional embedding with shape `(temporal_size * grid_size[0] * grid_size[1], embed_dim/2)`.
+    """
+    start, stop = crops_coords
+    grid_h = np.linspace(start[0], stop[0], grid_size[0], endpoint=False, dtype=np.float32)
+    grid_w = np.linspace(start[1], stop[1], grid_size[1], endpoint=False, dtype=np.float32)
+    grid_t = np.linspace(0, temporal_size, temporal_size, endpoint=False, dtype=np.float32)
+
+    # Compute dimensions for each axis
+    dim_t = embed_dim // 4
+    dim_h = embed_dim // 8 * 3
+    dim_w = embed_dim // 8 * 3
+
+    # Temporal frequencies
+    freqs_t = 1.0 / (theta ** (torch.arange(0, dim_t, 2).float() / dim_t))
+    grid_t = torch.from_numpy(grid_t).float()
+    freqs_t = torch.einsum("n , f -> n f", grid_t, freqs_t)
+    freqs_t = freqs_t.repeat_interleave(2, dim=-1)
+
+    # Spatial frequencies for height and width
+    freqs_h = 1.0 / (theta ** (torch.arange(0, dim_h, 2).float() / dim_h))
+    freqs_w = 1.0 / (theta ** (torch.arange(0, dim_w, 2).float() / dim_w))
+    grid_h = torch.from_numpy(grid_h).float()
+    grid_w = torch.from_numpy(grid_w).float()
+    freqs_h = torch.einsum("n , f -> n f", grid_h, freqs_h)
+    freqs_w = torch.einsum("n , f -> n f", grid_w, freqs_w)
+    freqs_h = freqs_h.repeat_interleave(2, dim=-1)
+    freqs_w = freqs_w.repeat_interleave(2, dim=-1)
+
+    # Broadcast and concatenate tensors along specified dimension
+    def broadcast(tensors, dim=-1):
+        num_tensors = len(tensors)
+        shape_lens = {len(t.shape) for t in tensors}
+        assert len(shape_lens) == 1, "tensors must all have the same number of dimensions"
+        shape_len = list(shape_lens)[0]
+        dim = (dim + shape_len) if dim < 0 else dim
+        dims = list(zip(*(list(t.shape) for t in tensors)))
+        expandable_dims = [(i, val) for i, val in enumerate(dims) if i != dim]
+        assert all(
+            [*(len(set(t[1])) <= 2 for t in expandable_dims)]
+        ), "invalid dimensions for broadcastable concatenation"
+        max_dims = [(t[0], max(t[1])) for t in expandable_dims]
+        expanded_dims = [(t[0], (t[1],) * num_tensors) for t in max_dims]
+        expanded_dims.insert(dim, (dim, dims[dim]))
+        expandable_shapes = list(zip(*(t[1] for t in expanded_dims)))
+        tensors = [t[0].expand(*t[1]) for t in zip(tensors, expandable_shapes)]
+        return torch.cat(tensors, dim=dim)
+
+    freqs = broadcast((freqs_t[:, None, None, :], freqs_h[None, :, None, :], freqs_w[None, None, :, :]), dim=-1)
+
+    t, h, w, d = freqs.shape
+    freqs = freqs.view(t * h * w, d)
+
+    # Generate sine and cosine components
+    sin = freqs.sin()
+    cos = freqs.cos()
+
+    if use_real:
+        return cos, sin
+    else:
+        freqs_cis = torch.polar(torch.ones_like(freqs), freqs)
+        return freqs_cis
+
+
 def get_2d_rotary_pos_embed(embed_dim, crops_coords, grid_size, use_real=True):
     """
     RoPE for image tokens with 2d structure.