File size: 12,705 Bytes
bc29ee3 | 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 | #!/usr/bin/env python3
"""Evaluate 2x/4x long FFFF and one-block Layer-17 FPPF rollouts."""
from __future__ import annotations
import argparse
import json
import os
import sys
import time
from pathlib import Path
def _preparse_gpu() -> str:
parser = argparse.ArgumentParser(add_help=False)
parser.add_argument("--gpu", default="0")
args, _ = parser.parse_known_args()
os.environ["CUDA_VISIBLE_DEVICES"] = str(args.gpu)
return str(args.gpu)
PHYSICAL_GPU = _preparse_gpu()
import lpips
import torch
from omegaconf import OmegaConf
from torchvision.io import write_video
REPO_ROOT = Path(__file__).resolve().parents[1]
if str(REPO_ROOT) not in sys.path:
sys.path.insert(0, str(REPO_ROOT))
from predictor_training.offline_data import TOKENS_PER_CHUNK
from scripts.evaluate_single_block_fppf import (
DEFAULT_PROMPT_IDS,
FRAMES_PER_CHUNK,
LATENT_CHANNELS,
LATENT_HEIGHT,
LATENT_WIDTH,
NUM_DENOISING_STEPS,
FinalHiddenCapture,
atomic_json,
build_pipeline,
discover_experiments,
frame_metrics,
load_predictor,
load_prompt_metadata,
pixels_to_u8,
predictor_step,
reset_kv_and_load_cross_cache,
)
from utils.misc import set_seed
from utils.wan_wrapper import WanVAEWrapper
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--gpu", default=PHYSICAL_GPU)
parser.add_argument("--prompt_ids", type=int, nargs="+", required=True)
parser.add_argument("--latent_lengths", type=int, nargs="+", default=[42, 84])
parser.add_argument(
"--config_path", type=Path, default=Path("configs/self_forcing_sid.yaml")
)
parser.add_argument(
"--checkpoint_path", type=Path, default=Path("checkpoints/self_forcing_dmd.pt")
)
parser.add_argument(
"--dataset_root", type=Path,
default=Path("outputs/predictor_offline_100_all_blocks"),
)
parser.add_argument(
"--sweep_dir", type=Path, default=Path("outputs/single_block_init_sweep")
)
parser.add_argument(
"--output_dir", type=Path, default=Path("outputs/long_video_2x4x_eval")
)
parser.add_argument("--metric_batch_size", type=int, default=4)
parser.add_argument("--generation_seed", type=int, default=0)
args = parser.parse_args()
if any(length <= 0 or length % FRAMES_PER_CHUNK for length in args.latent_lengths):
parser.error("Latent lengths must be positive multiples of 3")
if any(prompt not in DEFAULT_PROMPT_IDS for prompt in args.prompt_ids):
parser.error("This evaluation is restricted to validation prompt IDs 80..99")
return args
def resolve(path: Path) -> Path:
return path.resolve() if path.is_absolute() else (REPO_ROOT / path).resolve()
@torch.inference_mode()
def generate_rollout(
*, pipeline, dataset_root: Path, prompt_id: int, latent_length: int,
generation_seed: int, device: torch.device, predictor, source_layer: int | None,
schedule: str,
) -> tuple[torch.Tensor, dict[str, float | int]]:
if schedule not in {"FFFF", "FPPF"}:
raise ValueError(schedule)
if schedule == "FPPF" and (predictor is None or source_layer is None):
raise ValueError("FPPF requires the Predictor")
num_chunks = latent_length // FRAMES_PER_CHUNK
reset_kv_and_load_cross_cache(pipeline, dataset_root, prompt_id, device)
set_seed(generation_seed)
noise = torch.randn(
1, latent_length, LATENT_CHANNELS, LATENT_HEIGHT, LATENT_WIDTH,
dtype=torch.bfloat16, device=device,
)
teacher = pipeline.generator.model
text_dim = int(teacher.text_embedding[0].in_features)
conditional_dict = {
"prompt_embeds": torch.zeros(
1, 1, text_dim, dtype=torch.bfloat16, device=device
)
}
timesteps = pipeline.denoising_step_list.to(device=device)
output_chunks = []
previous_chunk_hidden = None
capture = FinalHiddenCapture(teacher)
full_calls = predictor_calls = 0
started = time.perf_counter()
try:
for chunk in range(num_chunks):
noisy_input = noise[:, chunk * 3:(chunk + 1) * 3]
current_hidden = [None] * NUM_DENOISING_STEPS
denoised_pred = timestep = None
for step, current_timestep in enumerate(timesteps):
timestep = torch.ones(
[1, FRAMES_PER_CHUNK], dtype=torch.int64, device=device
) * current_timestep
use_predictor = schedule == "FPPF" and chunk > 0 and step in {1, 2}
if use_predictor:
pred_hidden, flow, _ = predictor_step(
predictor=predictor,
teacher=teacher,
noisy_input=noisy_input,
timestep=timestep,
anchor_hidden=current_hidden[step - 1],
previous_hidden=previous_chunk_hidden[step],
history_cache=pipeline.kv_cache1[source_layer],
cross_cache=pipeline.crossattn_cache[source_layer],
current_start=chunk * TOKENS_PER_CHUNK,
)
denoised_pred = pipeline.generator._convert_flow_pred_to_x0(
flow_pred=flow.flatten(0, 1),
xt=noisy_input.flatten(0, 1),
timestep=timestep.flatten(0, 1),
).unflatten(0, flow.shape[:2])
current_hidden[step] = pred_hidden
predictor_calls += 1
else:
capture.start()
_, denoised_pred = pipeline.generator(
noisy_image_or_video=noisy_input,
conditional_dict=conditional_dict,
timestep=timestep,
kv_cache=pipeline.kv_cache1,
crossattn_cache=pipeline.crossattn_cache,
current_start=chunk * TOKENS_PER_CHUNK,
)
current_hidden[step] = capture.finish()
full_calls += 1
if step < NUM_DENOISING_STEPS - 1:
flat = denoised_pred.flatten(0, 1)
noisy_input = pipeline.scheduler.add_noise(
flat, torch.randn_like(flat),
timesteps[step + 1] * torch.ones(
[FRAMES_PER_CHUNK], dtype=torch.long, device=device
),
).unflatten(0, denoised_pred.shape[:2])
output_chunks.append(denoised_pred)
pipeline.generator(
noisy_image_or_video=denoised_pred,
conditional_dict=conditional_dict,
timestep=torch.ones_like(timestep) * pipeline.args.context_noise,
kv_cache=pipeline.kv_cache1,
crossattn_cache=pipeline.crossattn_cache,
current_start=chunk * TOKENS_PER_CHUNK,
)
previous_chunk_hidden = current_hidden
finally:
capture.close()
torch.cuda.synchronize()
return torch.cat(output_chunks, dim=1), {
"generation_time_s": time.perf_counter() - started,
"full_calls": full_calls,
"predictor_calls": predictor_calls,
"num_chunks": num_chunks,
}
def save_mp4(frames: torch.Tensor, path: Path) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
write_video(
str(path), frames.permute(0, 2, 3, 1), fps=16,
video_codec="libx264", options={"crf": "18"},
)
def main() -> None:
args = parse_args()
for field in ("config_path", "checkpoint_path", "dataset_root", "sweep_dir", "output_dir"):
setattr(args, field, resolve(getattr(args, field)))
args.output_dir.mkdir(parents=True, exist_ok=True)
atomic_json(args.output_dir / "manifest.json", {
"status": "running", "physical_gpu": str(args.gpu),
"prompt_ids": args.prompt_ids, "latent_lengths": args.latent_lengths,
"methods": ["FFFF", "FPPF_teacher_layer_17"],
"generation_seed": args.generation_seed,
})
device = torch.device("cuda")
torch.set_grad_enabled(False)
set_seed(args.generation_seed)
config = OmegaConf.merge(
OmegaConf.load(REPO_ROOT / "configs/default_config.yaml"),
OmegaConf.load(args.config_path),
)
# The released checkpoint uses full attention over its 21-latent training
# horizon. Long inference keeps exactly that horizon as a rolling window;
# within the first 21 latents this is numerically the same attention span.
config.model_kwargs.local_attn_size = 21
vae = WanVAEWrapper().to(device=device, dtype=torch.bfloat16).eval()
pipeline = build_pipeline(config, args.checkpoint_path, vae, device)
experiment = discover_experiments(
args.sweep_dir, ["teacher_layer_17"], None
)[0]
predictor = load_predictor(pipeline.generator.model, experiment, device)
lpips_model = lpips.LPIPS(net="alex", verbose=False).to(device).eval()
lpips_model.requires_grad_(False)
for prompt_id in args.prompt_ids:
prompt = load_prompt_metadata(args.dataset_root, prompt_id)["prompt"]
for latent_length in args.latent_lengths:
run_dir = args.output_dir / f"latent_{latent_length}" / f"prompt_{prompt_id:04d}"
result_path = run_dir / "metrics.json"
if result_path.exists():
existing = json.loads(result_path.read_text(encoding="utf-8"))
if existing.get("status") == "complete":
print(f"[skip] latent={latent_length} prompt={prompt_id}", flush=True)
continue
print(f"[run] latent={latent_length} prompt={prompt_id} FFFF", flush=True)
reference_latent, ffff_counts = generate_rollout(
pipeline=pipeline, dataset_root=args.dataset_root,
prompt_id=prompt_id, latent_length=latent_length,
generation_seed=args.generation_seed, device=device,
predictor=None, source_layer=None, schedule="FFFF",
)
with torch.autocast(device_type="cuda", dtype=torch.bfloat16):
reference_pixels = vae.decode_to_pixel(reference_latent, use_cache=False)
reference_u8 = pixels_to_u8(reference_pixels)
save_mp4(reference_u8, run_dir / "ffff.mp4")
del reference_latent, reference_pixels
if hasattr(vae.model, "clear_cache"):
vae.model.clear_cache()
torch.cuda.empty_cache()
print(f"[run] latent={latent_length} prompt={prompt_id} FPPF", flush=True)
prediction_latent, fppf_counts = generate_rollout(
pipeline=pipeline, dataset_root=args.dataset_root,
prompt_id=prompt_id, latent_length=latent_length,
generation_seed=args.generation_seed, device=device,
predictor=predictor, source_layer=17, schedule="FPPF",
)
with torch.autocast(device_type="cuda", dtype=torch.bfloat16):
prediction_pixels = vae.decode_to_pixel(prediction_latent, use_cache=False)
prediction_u8 = pixels_to_u8(prediction_pixels)
save_mp4(prediction_u8, run_dir / "fppf_layer17.mp4")
metrics = frame_metrics(
reference_u8=reference_u8, prediction_u8=prediction_u8,
lpips_model=lpips_model, batch_size=args.metric_batch_size,
device=device,
)
atomic_json(result_path, {
"status": "complete", "prompt_id": prompt_id, "prompt": prompt,
"latent_length": latent_length, "decoded_frames": metrics["num_frames"],
"reference": "FFFF same prompt/seed/noise",
"predictor": "single_block_teacher_layer_17",
"ffff": ffff_counts, "fppf": fppf_counts, **metrics,
})
print(
f"[result] latent={latent_length} prompt={prompt_id} "
f"psnr={metrics['psnr']:.4f} ssim={metrics['ssim']:.6f} "
f"lpips={metrics['lpips']:.6f}", flush=True,
)
del prediction_latent, prediction_pixels, reference_u8, prediction_u8
if hasattr(vae.model, "clear_cache"):
vae.model.clear_cache()
torch.cuda.empty_cache()
manifest = json.loads((args.output_dir / "manifest.json").read_text(encoding="utf-8"))
manifest["status"] = "complete"
atomic_json(args.output_dir / "manifest.json", manifest)
if __name__ == "__main__":
main()
|