File size: 11,874 Bytes
2bfd19c | 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 | # Copyright (c) 2025 SandAI. 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.
"""
MotionCache: motion-aware token-wise cache reuse for autoregressive video generation.
Implements the coarse-to-fine schedule from Xu et al. (2026):
- Phase 1 (first K steps after warmup): chunk-wise binary reuse (FlowCache-style)
- Phase 2: motion-weighted per-token accumulation and selective reuse
"""
import json
import os
from typing import Dict, Optional, Tuple
import torch
from einops import rearrange
from .cachereuse import ChunkWiseCache
class MotionWiseCache(ChunkWiseCache):
"""
Motion-aware cache extending chunk-wise FlowCache with token-level reuse.
Hyperparameters (paper Appendix C for MAGI-1):
alpha: soft-mapping floor for static tokens (default 0.5)
phase1_steps (K): chunk-wise phase duration before token-wise mode (default 9)
rel_l1_thresh (tau): accumulator threshold for token activation
warmup_steps (m): global steps with reuse disabled (default 5)
"""
def __init__(
self,
rel_l1_thresh: float = 0.015,
warmup_steps: int = 5,
phase1_steps: int = 9,
alpha: float = 0.5,
discard_nearly_clean_chunk: bool = False,
log: bool = False,
metric_stats_path: Optional[str] = None,
eps: float = 1e-8,
):
super().__init__(
rel_l1_thresh=rel_l1_thresh,
warmup_steps=warmup_steps,
discard_nearly_clean_chunk=discard_nearly_clean_chunk,
log=log,
metric_stats_path=metric_stats_path,
)
self.phase1_steps = phase1_steps
self.alpha = alpha
self.eps = eps
self.token_accumulator: Dict[int, torch.Tensor] = {}
self.token_active_mask: Dict[int, torch.Tensor] = {}
self.token_motion_weights: Dict[int, torch.Tensor] = {}
self.prev_latent_chunks: Dict[int, torch.Tensor] = {}
self.prev_chunk_last_frame: Dict[int, torch.Tensor] = {}
self.previous_velocity: Dict[int, torch.Tensor] = {}
self.chunk_sparse_flags: Dict[int, bool] = {}
def reset(self):
super().reset()
self.token_accumulator.clear()
self.token_active_mask.clear()
self.token_motion_weights.clear()
self.prev_latent_chunks.clear()
self.prev_chunk_last_frame.clear()
self.previous_velocity.clear()
self.chunk_sparse_flags.clear()
@staticmethod
def expand_token_mask_to_output(
token_mask: torch.Tensor,
output: torch.Tensor,
) -> torch.Tensor:
"""Expand [N, T, H, W] latent mask to match velocity/output [N, C, T, H, W]."""
return token_mask.unsqueeze(1).expand_as(output).to(dtype=output.dtype)
def in_phase1(self, chunk_id: int, chunk_denoise_count: Dict[int, int]) -> bool:
"""Return True while chunk i is still in coarse chunk-wise phase (denoise step < K)."""
return chunk_denoise_count.get(chunk_id, 0) < self.phase1_steps
def compute_motion_weights(
self,
x_chunk: torch.Tensor,
chunk_id: int,
chunk_offset: int,
) -> torch.Tensor:
"""
Compute motion-aware importance weights W in [alpha, 1] per latent frame.
Args:
x_chunk: Latent tensor [N, C, T, H, W] at current denoising step
chunk_id: Global chunk index
chunk_offset: Index of first generated chunk
Returns:
Weights tensor [N, T, H, W]
"""
_, _, num_frames, _, _ = x_chunk.shape
device = x_chunk.device
dtype = x_chunk.dtype
importance = torch.zeros(
x_chunk.size(0), num_frames, x_chunk.size(3), x_chunk.size(4),
device=device, dtype=dtype,
)
for frame_idx in range(num_frames):
if frame_idx > 0:
diff = (x_chunk[:, :, frame_idx] - x_chunk[:, :, frame_idx - 1]).abs().sum(dim=1)
elif chunk_id > chunk_offset:
prev_frame = self.prev_chunk_last_frame.get(chunk_id - 1)
if prev_frame is not None:
diff = (x_chunk[:, :, 0] - prev_frame).abs().sum(dim=1)
else:
diff = torch.zeros(
x_chunk.size(0), x_chunk.size(3), x_chunk.size(4),
device=device, dtype=dtype,
)
else:
continue
importance[:, frame_idx] = diff
if chunk_id == chunk_offset and num_frames > 1:
importance[:, 0] = importance[:, 1]
weights = torch.zeros_like(importance)
for frame_idx in range(num_frames):
frame_importance = importance[:, frame_idx]
min_val = frame_importance.amin(dim=(1, 2), keepdim=True)
max_val = frame_importance.amax(dim=(1, 2), keepdim=True)
normalized = (frame_importance - min_val) / (max_val - min_val + self.eps)
weights[:, frame_idx] = self.alpha + (1.0 - self.alpha) * normalized
return weights
def compute_chunk_delta_l1(
self,
current_features: torch.Tensor,
prev_features: torch.Tensor,
) -> float:
"""Relative L1 distance between consecutive embedded features (Eq. 11)."""
diff = (current_features - prev_features).abs().mean()
denom = prev_features.abs().mean() + self.eps
return (diff / denom).item()
def update_token_policy(
self,
chunk_id: int,
x_chunk: torch.Tensor,
current_features: torch.Tensor,
chunk_offset: int,
chunk_denoise_count: Optional[Dict[int, int]] = None,
) -> torch.Tensor:
"""
Phase-2 token policy: update accumulators and return active mask.
Returns:
Boolean mask [N, T, H, W], True = compute, False = reuse cache
"""
if (
chunk_denoise_count is not None
and chunk_denoise_count.get(chunk_id, 0) == self.phase1_steps
):
mask = torch.ones(
x_chunk.size(0), x_chunk.size(2), x_chunk.size(3), x_chunk.size(4),
device=x_chunk.device, dtype=torch.bool,
)
self.token_active_mask[chunk_id] = mask
self.token_accumulator[chunk_id] = torch.zeros(
x_chunk.size(0), x_chunk.size(2), x_chunk.size(3), x_chunk.size(4),
device=x_chunk.device,
dtype=x_chunk.dtype,
)
return mask
prev_features = self.prev_metric_chunks.get(chunk_id)
if prev_features is None:
mask = torch.ones(
x_chunk.size(0), x_chunk.size(2), x_chunk.size(3), x_chunk.size(4),
device=x_chunk.device, dtype=torch.bool,
)
self.token_active_mask[chunk_id] = mask
return mask
delta_chunk = self.compute_chunk_delta_l1(current_features, prev_features)
weights = self.compute_motion_weights(x_chunk, chunk_id, chunk_offset)
self.token_motion_weights[chunk_id] = weights
if chunk_id not in self.token_accumulator:
self.token_accumulator[chunk_id] = torch.zeros_like(weights)
self.token_accumulator[chunk_id] = (
self.token_accumulator[chunk_id] + weights * delta_chunk
)
mask = self.token_accumulator[chunk_id] > self.rel_l1_thresh
self.token_active_mask[chunk_id] = mask
return mask
def reset_token_accumulator(self, chunk_id: int, mask: torch.Tensor):
"""Reset accumulator for tokens selected for computation."""
if chunk_id in self.token_accumulator:
self.token_accumulator[chunk_id] = torch.where(
mask, torch.zeros_like(self.token_accumulator[chunk_id]),
self.token_accumulator[chunk_id],
)
def should_skip_chunk_forward(
self,
chunk_id: int,
chunk_denoise_count: Dict[int, int],
) -> bool:
"""Return True if the entire chunk can skip the DiT forward pass."""
if self.in_phase1(chunk_id, chunk_denoise_count):
return self.chunk_reuse_flags.get(chunk_id, False)
mask = self.token_active_mask.get(chunk_id)
if mask is None:
return False
return not mask.any()
def store_latent_chunk(self, chunk_id: int, x_chunk: torch.Tensor):
"""Store latent for cross-chunk motion reference."""
self.prev_latent_chunks[chunk_id] = x_chunk.detach().clone()
self.prev_chunk_last_frame[chunk_id] = x_chunk[:, :, -1].detach().clone()
def get_token_mask(
self,
chunk_id: int,
chunk_denoise_count: Dict[int, int],
) -> Optional[torch.Tensor]:
if self.in_phase1(chunk_id, chunk_denoise_count):
return None
return self.token_active_mask.get(chunk_id)
def record_motion_decision(
self,
chunk_id: int,
reused: bool,
active_ratio: Optional[float] = None,
**kwargs,
):
if not self.metric_stats_path:
return
record = {
"infer_idx": kwargs.get("infer_idx"),
"cur_denoise_step": kwargs.get("cur_denoise_step"),
"denoise_stage": kwargs.get("denoise_stage"),
"denoise_idx": kwargs.get("denoise_idx"),
"chunk_idx": chunk_id,
"generated_chunk_idx": chunk_id - kwargs.get("chunk_offset", 0),
"chunk_denoise_count": kwargs.get("chunk_denoise_count_value"),
"phase": (
"phase1_chunk"
if self.in_phase1(chunk_id, kwargs.get("chunk_denoise_count", {}))
else "phase2_token"
),
"reused": bool(reused),
"execution": "reuse" if reused else "compute",
"active_token_ratio": active_ratio,
"phase1_steps": self.phase1_steps,
"alpha": self.alpha,
"rel_l1_thresh": self.rel_l1_thresh,
}
self.execution_records.append(record)
def save_metric_stats(self):
if not self.metric_stats_path:
return
save_dir = os.path.dirname(self.metric_stats_path)
if save_dir:
os.makedirs(save_dir, exist_ok=True)
payload = {
"description": (
"MotionCache metric stats. Phase 1 uses chunk-wise FlowCache policy for "
f"the first {self.phase1_steps} denoise steps per chunk; Phase 2 uses motion-weighted "
"token accumulation with alpha floor and rel_l1_thresh."
),
"hyperparameters": {
"alpha": self.alpha,
"phase1_steps": self.phase1_steps,
"warmup_steps": self.warmup_steps,
"rel_l1_thresh": self.rel_l1_thresh,
},
"chunk_execution_summary": self.get_execution_summary(),
"execution_records": self.execution_records,
"records": self.metric_records,
}
if self.metric_stats_path.endswith((".pt", ".pth")):
torch.save(payload, self.metric_stats_path)
else:
with open(self.metric_stats_path, "w") as f:
json.dump(payload, f, indent=2)
print(f"Saved MotionCache metric stats to {self.metric_stats_path}")
|