File size: 11,337 Bytes
ff5e4ca |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 |
# Copyright 2025 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 Any, List, Tuple
import torch
from diffusers.configuration_utils import FrozenDict
from diffusers.guiders import ClassifierFreeGuidance
from diffusers.models import AutoModel
from diffusers.schedulers import UniPCMultistepScheduler
from diffusers.utils import logging
from diffusers.utils.torch_utils import randn_tensor
from diffusers.modular_pipelines import (
BlockState,
LoopSequentialPipelineBlocks,
ModularPipelineBlocks,
PipelineState,
ModularPipeline,
)
from diffusers.modular_pipelines.modular_pipeline_utils import (
ComponentSpec,
InputParam,
OutputParam,
)
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
class WanRTStreamingLoopDenoiser(ModularPipelineBlocks):
model_name = "WanRTStreaming"
@property
def expected_components(self) -> List[ComponentSpec]:
return [ComponentSpec("transformer", AutoModel)]
@property
def description(self) -> str:
return (
"Step within the denoising loop that denoise the latents with guidance. "
"This block should be used to compose the `sub_blocks` attribute of a `LoopSequentialPipelineBlocks` "
"object (e.g. `WanRTStreamingDenoiseLoopWrapper`)"
)
@property
def inputs(self) -> List[Tuple[str, Any]]:
return [
InputParam("attention_kwargs"),
InputParam("block_idx"),
InputParam(
"latents",
required=True,
type_hint=torch.Tensor,
description="The initial latents to use for the denoising process. Can be generated in prepare_latent step.",
),
InputParam(
"prompt_embeds",
required=True,
type_hint=torch.Tensor,
),
InputParam(
"kv_cache",
required=True,
type_hint=torch.Tensor,
),
InputParam(
"crossattn_cache",
required=True,
type_hint=torch.Tensor,
),
InputParam(
"current_start_frame",
required=True,
type_hint=torch.Tensor,
),
InputParam(
"num_inference_steps",
required=True,
type_hint=int,
default=4,
description="The number of inference steps to use for the denoising process. Can be generated in set_timesteps step.",
),
InputParam(
kwargs_type="guider_input_fields",
description=(
"All conditional model inputs that need to be prepared with guider. "
"It should contain prompt_embeds/negative_prompt_embeds. "
"Please add `kwargs_type=guider_input_fields` to their parameter spec (`OutputParam`) when they are created and added to the pipeline state"
),
),
]
@torch.no_grad()
def __call__(
self,
components: ModularPipeline,
block_state: BlockState,
i: int,
t: torch.Tensor,
) -> PipelineState:
start_frame = min(
block_state.current_start_frame, components.config.kv_cache_num_frames
)
block_state.noise_pred = components.transformer(
x=block_state.latents,
t=t.expand(block_state.latents.shape[0], block_state.num_frames_per_block),
context=block_state.prompt_embeds,
kv_cache=block_state.kv_cache,
seq_len=components.config.seq_length,
crossattn_cache=block_state.crossattn_cache,
current_start=start_frame * components.config.frame_seq_length,
cache_start=start_frame * components.config.frame_seq_length,
)
return components, block_state
class WanRTStreamingLoopAfterDenoiser(ModularPipelineBlocks):
model_name = "WanRTStreaming"
@property
def expected_components(self) -> List[ComponentSpec]:
return [
ComponentSpec("scheduler", UniPCMultistepScheduler),
]
@property
def description(self) -> str:
return (
"step within the denoising loop that update the latents. "
"This block should be used to compose the `sub_blocks` attribute of a `LoopSequentialPipelineBlocks` "
"object (e.g. `WanRTStreamingDenoiseLoopWrapper`)"
)
@property
def inputs(self) -> List[Tuple[str, Any]]:
return []
@property
def intermediate_inputs(self) -> List[str]:
return [
InputParam("generator"),
InputParam("block_id"),
]
@property
def intermediate_outputs(self) -> List[OutputParam]:
return [
OutputParam(
"latents", type_hint=torch.Tensor, description="The denoised latents"
)
]
@torch.no_grad()
def __call__(
self,
components: ModularPipeline,
block_state: BlockState,
i: int,
t: torch.Tensor,
):
# Perform scheduler step using the predicted output
latents_dtype = block_state.latents.dtype
timesteps = block_state.all_timesteps
sigmas = block_state.sigmas
timestep_id = torch.argmin((timesteps - t).abs())
sigma_t = sigmas[timestep_id]
# Perform computation in double precision, then convert back once
latents = (
block_state.latents.double()
- sigma_t.double() * block_state.noise_pred.double()
).to(latents_dtype)
block_state.latents = latents
return components, block_state
class WanRTStreamingDenoiseLoopWrapper(LoopSequentialPipelineBlocks):
model_name = "WanRTStreaming"
@property
def description(self) -> str:
return (
"Streaming denoising loop that processes a single block with persistent KV cache. "
"Recomputes cache from context frames, denoises current block, and updates cache."
)
def add_noise(self, components, block_state, sample, noise, timestep, index):
timesteps = block_state.all_timesteps
sigmas = block_state.sigmas.to(timesteps.device)
if timestep.ndim == 2:
timestep = timestep.flatten(0, 1)
timestep_id = torch.argmin(
(timesteps.unsqueeze(0) - timestep.unsqueeze(1)).abs(), dim=1
)
sigma = sigmas[timestep_id].reshape(-1, 1, 1, 1)
sample = (
1 - sigma.double()
) * sample.double() + sigma.double() * noise.double()
sample = sample.type_as(noise)
return sample
@property
def loop_inputs(self) -> List[InputParam]:
return [
InputParam(
"timesteps",
required=True,
type_hint=torch.Tensor,
description="The timesteps to use for the denoising process. Can be generated in set_timesteps step.",
),
InputParam(
"all_timesteps",
required=True,
type_hint=torch.Tensor,
description="The timesteps to use for the denoising process. Can be generated in set_timesteps step.",
),
InputParam(
"sigmas",
required=True,
type_hint=torch.Tensor,
description="The timesteps to use for the denoising process. Can be generated in set_timesteps step.",
),
InputParam("final_latents", type_hint=torch.Tensor),
InputParam(
"num_inference_steps",
required=True,
type_hint=int,
description="The number of inference steps to use for the denoising process. Can be generated in set_timesteps step.",
),
InputParam(
"num_frames_per_block",
required=True,
type_hint=int,
default=3,
),
InputParam(
"current_start_frame",
required=True,
type_hint=int,
),
InputParam(
"block_idx",
),
InputParam(
"generator",
),
]
@torch.no_grad()
def __call__(
self, components: ModularPipeline, state: PipelineState
) -> PipelineState:
block_state = self.get_block_state(state)
for i, t in enumerate(block_state.timesteps):
components, block_state = self.loop_step(components, block_state, i=i, t=t)
if i < (block_state.num_inference_steps - 1):
t1 = block_state.timesteps[i + 1]
block_state.latents = (
self.add_noise(
components,
block_state,
block_state.latents.transpose(1, 2).squeeze(0),
randn_tensor(
block_state.latents.transpose(1, 2).squeeze(0).shape,
device=block_state.latents.device,
dtype=block_state.latents.dtype,
generator=block_state.generator,
),
t1.expand(
block_state.latents.shape[0],
block_state.num_frames_per_block,
),
i,
)
.unsqueeze(0)
.transpose(1, 2)
)
# Update the state
block_state.final_latents[
:,
:,
block_state.current_start_frame : block_state.current_start_frame
+ block_state.num_frames_per_block,
] = block_state.latents
self.set_block_state(state, block_state)
return components, state
class WanRTStreamingDenoiseStep(WanRTStreamingDenoiseLoopWrapper):
block_classes = [
WanRTStreamingLoopDenoiser,
WanRTStreamingLoopAfterDenoiser,
]
block_names = ["denoiser", "after_denoiser"]
@property
def description(self) -> str:
return (
"Denoise step that iteratively denoise the latents. \n"
"Its loop logic is defined in `WanRTStreamingDenoiseLoopWrapper.__call__` method \n"
"At each iteration, it runs blocks defined in `sub_blocks` sequencially:\n"
" - `WanRTStreamingLoopDenoiser`\n"
" - `WanRTStreamingLoopAfterDenoiser`\n"
"This block supports both text2vid tasks."
)
|