mirror of
https://github.com/Wan-Video/Wan2.1.git
synced 2025-11-04 14:16:57 +00:00
424 lines
19 KiB
Python
424 lines
19 KiB
Python
# borrowed from svg-project/Sparse-VideoGen
|
|
|
|
from typing import Any, Dict, Optional, Tuple, Union
|
|
|
|
import torch
|
|
|
|
from diffusers.models.transformers.transformer_wan import WanTransformerBlock, WanTransformer3DModel
|
|
from diffusers import WanPipeline
|
|
from diffusers.models.modeling_outputs import Transformer2DModelOutput
|
|
from diffusers.utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
|
|
logger = logging.get_logger(__name__)
|
|
from typing import Any, Callable, Dict, List, Optional, Union
|
|
from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
|
|
from diffusers.pipelines.wan.pipeline_output import WanPipelineOutput
|
|
import torch.distributed as dist
|
|
|
|
try:
|
|
from xfuser.core.distributed import (
|
|
get_ulysses_parallel_world_size,
|
|
get_ulysses_parallel_rank,
|
|
get_sp_group
|
|
)
|
|
except:
|
|
pass
|
|
|
|
class WanTransformerBlock_Sparse(WanTransformerBlock):
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
encoder_hidden_states: torch.Tensor,
|
|
temb: torch.Tensor,
|
|
rotary_emb: torch.Tensor,
|
|
numeral_timestep: Optional[int] = None,
|
|
) -> torch.Tensor:
|
|
if temb.ndim == 4:
|
|
# temb: batch_size, seq_len, 6, inner_dim (wan2.2 ti2v)
|
|
shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = (
|
|
self.scale_shift_table.unsqueeze(0) + temb.float()
|
|
).chunk(6, dim=2)
|
|
# batch_size, seq_len, 1, inner_dim
|
|
shift_msa = shift_msa.squeeze(2)
|
|
scale_msa = scale_msa.squeeze(2)
|
|
gate_msa = gate_msa.squeeze(2)
|
|
c_shift_msa = c_shift_msa.squeeze(2)
|
|
c_scale_msa = c_scale_msa.squeeze(2)
|
|
c_gate_msa = c_gate_msa.squeeze(2)
|
|
else:
|
|
# temb: batch_size, 6, inner_dim (wan2.1/wan2.2 14B)
|
|
shift_msa, scale_msa, gate_msa, c_shift_msa, c_scale_msa, c_gate_msa = (
|
|
self.scale_shift_table + temb.float()
|
|
).chunk(6, dim=1)
|
|
|
|
# 1. Self-attention
|
|
norm_hidden_states = (self.norm1(hidden_states.float()) * (1 + scale_msa) + shift_msa).type_as(hidden_states)
|
|
attn_output = self.attn1(hidden_states=norm_hidden_states, rotary_emb=rotary_emb, numerical_timestep=numeral_timestep)
|
|
hidden_states = (hidden_states.float() + attn_output * gate_msa).type_as(hidden_states).contiguous()
|
|
|
|
# 2. Cross-attention
|
|
norm_hidden_states = self.norm2(hidden_states.float()).type_as(hidden_states)
|
|
attn_output = self.attn2(hidden_states=norm_hidden_states, encoder_hidden_states=encoder_hidden_states)
|
|
hidden_states = hidden_states + attn_output
|
|
|
|
# 3. Feed-forward
|
|
norm_hidden_states = (self.norm3(hidden_states.float()) * (1 + c_scale_msa) + c_shift_msa).type_as(
|
|
hidden_states
|
|
)
|
|
ff_output = self.ffn(norm_hidden_states)
|
|
hidden_states = (hidden_states.float() + ff_output.float() * c_gate_msa).type_as(hidden_states)
|
|
|
|
return hidden_states
|
|
|
|
class WanTransformer3DModel_Sparse(WanTransformer3DModel):
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
timestep: torch.LongTensor,
|
|
encoder_hidden_states: torch.Tensor,
|
|
numeral_timestep: Optional[int] = None,
|
|
encoder_hidden_states_image: Optional[torch.Tensor] = None,
|
|
return_dict: bool = True,
|
|
attention_kwargs: Optional[Dict[str, Any]] = None,
|
|
) -> Union[torch.Tensor, Dict[str, torch.Tensor]]:
|
|
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."
|
|
)
|
|
|
|
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)
|
|
|
|
# timestep shape: batch_size, or batch_size, seq_len (wan 2.2 ti2v)
|
|
if timestep.ndim == 2:
|
|
ts_seq_len = timestep.shape[1]
|
|
timestep = timestep.flatten() # batch_size * seq_len
|
|
else:
|
|
ts_seq_len = None
|
|
|
|
temb, timestep_proj, encoder_hidden_states, encoder_hidden_states_image = self.condition_embedder(
|
|
timestep, encoder_hidden_states, encoder_hidden_states_image, timestep_seq_len=ts_seq_len
|
|
)
|
|
|
|
if ts_seq_len is not None:
|
|
# batch_size, seq_len, 6, inner_dim
|
|
timestep_proj = timestep_proj.unflatten(2, (6, -1))
|
|
else:
|
|
# batch_size, 6, inner_dim
|
|
timestep_proj = timestep_proj.unflatten(1, (6, -1))
|
|
|
|
if encoder_hidden_states_image is not None:
|
|
encoder_hidden_states = torch.concat([encoder_hidden_states_image, encoder_hidden_states], dim=1)
|
|
|
|
if dist.is_initialized() and get_ulysses_parallel_world_size() > 1:
|
|
# split video latents on dim TS
|
|
hidden_states = torch.chunk(hidden_states, get_ulysses_parallel_world_size(), dim=-2)[get_ulysses_parallel_rank()]
|
|
rotary_emb = (
|
|
torch.chunk(rotary_emb[0], get_ulysses_parallel_world_size(), dim=1)[get_ulysses_parallel_rank()],
|
|
torch.chunk(rotary_emb[1], get_ulysses_parallel_world_size(), dim=1)[get_ulysses_parallel_rank()],
|
|
)
|
|
|
|
# 4. Transformer blocks
|
|
if torch.is_grad_enabled() and self.gradient_checkpointing:
|
|
for block in self.blocks:
|
|
hidden_states = self._gradient_checkpointing_func(
|
|
block, hidden_states, encoder_hidden_states, timestep_proj, rotary_emb, numeral_timestep=numeral_timestep
|
|
)
|
|
else:
|
|
for block in self.blocks:
|
|
hidden_states = block(
|
|
hidden_states,
|
|
encoder_hidden_states,
|
|
timestep_proj,
|
|
rotary_emb,
|
|
numeral_timestep=numeral_timestep,
|
|
)
|
|
|
|
# 5. Output norm, projection & unpatchify
|
|
if temb.ndim == 3:
|
|
# batch_size, seq_len, inner_dim (wan 2.2 ti2v)
|
|
shift, scale = (self.scale_shift_table.unsqueeze(0) + temb.unsqueeze(2)).chunk(2, dim=2)
|
|
shift = shift.squeeze(2)
|
|
scale = scale.squeeze(2)
|
|
else:
|
|
# batch_size, inner_dim
|
|
shift, scale = (self.scale_shift_table + temb.unsqueeze(1)).chunk(2, dim=1)
|
|
|
|
# Move the shift and scale tensors to the same device as hidden_states.
|
|
# When using multi-GPU inference via accelerate these will be on the
|
|
# first device rather than the last device, which hidden_states ends up
|
|
# on.
|
|
shift = shift.to(hidden_states.device)
|
|
scale = scale.to(hidden_states.device)
|
|
|
|
hidden_states = (self.norm_out(hidden_states.float()) * (1 + scale) + shift).type_as(hidden_states)
|
|
hidden_states = self.proj_out(hidden_states)
|
|
|
|
if dist.is_initialized() and get_ulysses_parallel_world_size() > 1:
|
|
hidden_states = get_sp_group().all_gather(hidden_states, dim=-2)
|
|
|
|
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,)
|
|
|
|
return Transformer2DModelOutput(sample=output)
|
|
|
|
class WanPipeline_Sparse(WanPipeline):
|
|
@torch.no_grad()
|
|
def __call__(
|
|
self,
|
|
prompt: Union[str, List[str]] = None,
|
|
negative_prompt: Union[str, List[str]] = None,
|
|
height: int = 480,
|
|
width: int = 832,
|
|
num_frames: int = 81,
|
|
num_inference_steps: int = 50,
|
|
guidance_scale: float = 5.0,
|
|
num_videos_per_prompt: Optional[int] = 1,
|
|
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
|
latents: Optional[torch.Tensor] = None,
|
|
prompt_embeds: Optional[torch.Tensor] = None,
|
|
negative_prompt_embeds: Optional[torch.Tensor] = None,
|
|
output_type: Optional[str] = "np",
|
|
return_dict: bool = True,
|
|
attention_kwargs: Optional[Dict[str, Any]] = None,
|
|
callback_on_step_end: Optional[
|
|
Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
|
|
] = None,
|
|
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
|
|
max_sequence_length: int = 512,
|
|
):
|
|
r"""
|
|
The call function to 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`, defaults to `480`):
|
|
The height in pixels of the generated image.
|
|
width (`int`, defaults to `832`):
|
|
The width in pixels of the generated image.
|
|
num_frames (`int`, defaults to `81`):
|
|
The number of frames in the generated video.
|
|
num_inference_steps (`int`, defaults to `50`):
|
|
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
|
|
expense of slower inference.
|
|
guidance_scale (`float`, defaults to `5.0`):
|
|
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*):
|
|
A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
|
|
generation deterministic.
|
|
latents (`torch.Tensor`, *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 is generated by sampling using the supplied random `generator`.
|
|
prompt_embeds (`torch.Tensor`, *optional*):
|
|
Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
|
|
provided, text embeddings are generated from the `prompt` input argument.
|
|
output_type (`str`, *optional*, defaults to `"pil"`):
|
|
The output format of the generated image. Choose between `PIL.Image` or `np.array`.
|
|
return_dict (`bool`, *optional*, defaults to `True`):
|
|
Whether or not to return a [`WanPipelineOutput`] instead of a plain tuple.
|
|
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`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
|
|
A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
|
|
each denoising step during the inference. 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.
|
|
autocast_dtype (`torch.dtype`, *optional*, defaults to `torch.bfloat16`):
|
|
The dtype to use for the torch.amp.autocast.
|
|
|
|
Examples:
|
|
|
|
Returns:
|
|
[`~WanPipelineOutput`] or `tuple`:
|
|
If `return_dict` is `True`, [`WanPipelineOutput`] is returned, otherwise a `tuple` is returned where
|
|
the first element is a list with the generated images and the second element is a list of `bool`s
|
|
indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content.
|
|
"""
|
|
|
|
if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
|
|
callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
|
|
|
|
# 1. Check inputs. Raise error if not correct
|
|
self.check_inputs(
|
|
prompt,
|
|
negative_prompt,
|
|
height,
|
|
width,
|
|
prompt_embeds,
|
|
negative_prompt_embeds,
|
|
callback_on_step_end_tensor_inputs,
|
|
)
|
|
|
|
self._guidance_scale = guidance_scale
|
|
self._attention_kwargs = attention_kwargs
|
|
self._current_timestep = None
|
|
self._interrupt = False
|
|
|
|
device = self._execution_device
|
|
|
|
# 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]
|
|
|
|
# 3. Encode input prompt
|
|
prompt_embeds, negative_prompt_embeds = self.encode_prompt(
|
|
prompt=prompt,
|
|
negative_prompt=negative_prompt,
|
|
do_classifier_free_guidance=self.do_classifier_free_guidance,
|
|
num_videos_per_prompt=num_videos_per_prompt,
|
|
prompt_embeds=prompt_embeds,
|
|
negative_prompt_embeds=negative_prompt_embeds,
|
|
max_sequence_length=max_sequence_length,
|
|
device=device,
|
|
)
|
|
|
|
transformer_dtype = self.transformer.dtype
|
|
prompt_embeds = prompt_embeds.to(transformer_dtype)
|
|
if negative_prompt_embeds is not None:
|
|
negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype)
|
|
|
|
# 4. Prepare timesteps
|
|
self.scheduler.set_timesteps(num_inference_steps, device=device)
|
|
timesteps = self.scheduler.timesteps
|
|
|
|
# 5. 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,
|
|
torch.float32,
|
|
device,
|
|
generator,
|
|
latents,
|
|
)
|
|
|
|
# 6. Denoising loop
|
|
num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
|
|
self._num_timesteps = len(timesteps)
|
|
|
|
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
|
for i, t in enumerate(timesteps):
|
|
if self.interrupt:
|
|
continue
|
|
|
|
self._current_timestep = t
|
|
latent_model_input = latents.to(transformer_dtype)
|
|
timestep = t.expand(latents.shape[0])
|
|
|
|
noise_pred = self.transformer(
|
|
hidden_states=latent_model_input,
|
|
timestep=timestep,
|
|
encoder_hidden_states=prompt_embeds,
|
|
attention_kwargs=attention_kwargs,
|
|
return_dict=False,
|
|
numeral_timestep=i,
|
|
)[0]
|
|
|
|
if self.do_classifier_free_guidance:
|
|
noise_uncond = self.transformer(
|
|
hidden_states=latent_model_input,
|
|
timestep=timestep,
|
|
encoder_hidden_states=negative_prompt_embeds,
|
|
attention_kwargs=attention_kwargs,
|
|
return_dict=False,
|
|
numeral_timestep=i,
|
|
)[0]
|
|
noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond)
|
|
|
|
# compute the previous noisy sample x_t -> x_t-1
|
|
latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
|
|
|
|
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)
|
|
negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_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()
|
|
|
|
|
|
self._current_timestep = None
|
|
|
|
if not output_type == "latent":
|
|
latents = latents.to(self.vae.dtype)
|
|
latents_mean = (
|
|
torch.tensor(self.vae.config.latents_mean)
|
|
.view(1, self.vae.config.z_dim, 1, 1, 1)
|
|
.to(latents.device, latents.dtype)
|
|
)
|
|
latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
|
|
latents.device, latents.dtype
|
|
)
|
|
latents = latents / latents_std + latents_mean
|
|
video = self.vae.decode(latents, return_dict=False)[0]
|
|
video = self.video_processor.postprocess_video(video, output_type=output_type)
|
|
else:
|
|
video = latents
|
|
|
|
# Offload all models
|
|
self.maybe_free_model_hooks()
|
|
|
|
if not return_dict:
|
|
return (video,)
|
|
|
|
return WanPipelineOutput(frames=video)
|
|
|
|
def replace_sparse_forward():
|
|
WanTransformerBlock.forward = WanTransformerBlock_Sparse.forward
|
|
WanTransformer3DModel.forward = WanTransformer3DModel_Sparse.forward
|
|
WanPipeline.__call__ = WanPipeline_Sparse.__call__ |