File size: 16,344 Bytes
c6535db | 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 | import torch
from ..extrapolation import extrapolate_epsilon_linear, extrapolate_epsilon_richardson, extrapolate_epsilon_h4
from ...comfy_copy.res4lyf_sampling import get_res4lyf_step_with_model
from ..noise import get_eps_step_official
from ..skip import should_skip_model_call, validate_epsilon_hat, decide_skip_adaptive
from ..log import print_step_diag
def sample_step_gradient_estimation(model, noisy_latent, sigma_current, sigma_next, sigma_previous, s_in, extra_args,
epsilon_history, learning_ratio, smoothing_beta, predictor_type,
step_index, total_steps, add_noise_ratio=0.0, add_noise_type="whitened", skip_mode="none", skip_stats=None, debug=False, protect_last_steps=4, protect_first_steps=2, anchor_interval=None, max_consecutive_skips=None, official_comfy=False,
explicit_skip_indices=None, explicit_predictor=None, ge_gamma: float = 2.0):
x = noisy_latent
# Ensure commonly logged metrics are always defined
x_rms = None
if skip_stats is not None:
skip_stats["total_steps"] = skip_stats.get("total_steps", 0) + 1
# Final step guard: land on denoised
sigma_next_value = sigma_next.item() if torch.is_tensor(sigma_next) else float(sigma_next)
if abs(sigma_next_value) <= 1e-12:
den = model(x, sigma_current * s_in, **extra_args)
x = den
eps_real = den - noisy_latent
epsilon_history.append(eps_real)
if skip_stats is not None:
skip_stats["model_calls"] = skip_stats.get("model_calls", 0) + 1
skip_stats["consecutive_skips"] = 0
skip_stats["last_anchor_step"] = step_index
if len(epsilon_history) >= 3:
if predictor_type == "h4":
epsilon_hat = extrapolate_epsilon_h4(epsilon_history)
elif predictor_type == "richardson":
epsilon_hat = extrapolate_epsilon_richardson(epsilon_history)
else:
epsilon_hat = extrapolate_epsilon_linear(epsilon_history)
if epsilon_hat is not None:
learn_obs = (torch.norm(epsilon_hat) / (torch.norm(eps_real) + 1e-8)).item()
learning_ratio = smoothing_beta * learning_ratio + (1.0 - smoothing_beta) * learn_obs
learning_ratio = max(0.5, min(2.0, learning_ratio))
if debug:
print(f"gradient_est step {step_index} [LEARN]: learn_obs={learn_obs:.4f}, L={learning_ratio:.4f}, beta={smoothing_beta}")
return x, learning_ratio
# Target sigma and noise planning
target_sigma = sigma_next
sigma_up = None
alpha_ratio = None
if add_noise_ratio > 0.0 and float(sigma_next) > 0.0:
if official_comfy:
sigma_up, sigma_down = get_eps_step_official(sigma_current, sigma_next, eta=add_noise_ratio)
target_sigma = sigma_down
alpha_ratio = None
else:
sigma_up, _s, sigma_down, alpha_ratio = get_res4lyf_step_with_model(
model, sigma_current, sigma_next, add_noise_ratio, "hard"
)
target_sigma = sigma_down
dt = target_sigma - sigma_current
# d_prev from last REAL epsilon if available
d_prev = None
if sigma_previous is not None and len(epsilon_history) >= 1:
d_prev = -(epsilon_history[-1]) / sigma_previous
# Explicit skip indices take precedence
if explicit_skip_indices is not None and isinstance(explicit_skip_indices, set) and step_index in explicit_skip_indices:
es = skip_stats.get("explicit_streak", False) if skip_stats is not None else False
nl = skip_stats.get("needed_learns", 2) if skip_stats is not None else 2
allowed_by_streak = es or (nl <= 0)
if allowed_by_streak and len(epsilon_history) >= 2:
pred = (explicit_predictor or "linear")
if pred == "h4" and len(epsilon_history) >= 4:
epsilon_hat = extrapolate_epsilon_h4(epsilon_history)
tag = "explicit-h4"
elif (pred in ("richardson", "h3")) and len(epsilon_history) >= 3:
epsilon_hat = extrapolate_epsilon_richardson(epsilon_history)
tag = "explicit-h3"
else:
epsilon_hat = extrapolate_epsilon_linear(epsilon_history)
tag = "explicit-h2"
prev_eps = epsilon_history[-1] if len(epsilon_history) >= 1 else None
ok, reason, hat_norm, prev_norm = validate_epsilon_hat(epsilon_hat, prev_eps)
if ok:
if len(epsilon_history) >= 3:
epsilon_hat = epsilon_hat / max(learning_ratio, 1e-8)
d_hat = -(epsilon_hat) / sigma_current
dbar_hat = (ge_gamma - 1.0) * (d_hat - d_prev) if d_prev is not None else 0.0
# Clamp correction magnitude relative to base slope
if isinstance(dbar_hat, torch.Tensor):
try:
_ratio = float(torch.norm(dbar_hat) / (torch.norm(d_hat) + 1e-8))
except Exception:
_ratio = 0.0
if _ratio > 0.25:
dbar_hat = dbar_hat * (0.25 / _ratio)
x = x + (d_hat + (dbar_hat if isinstance(dbar_hat, torch.Tensor) else 0.0)) * dt
if skip_stats is not None:
skip_stats["skipped"] = skip_stats.get("skipped", 0) + 1
skip_stats["consecutive_skips"] = skip_stats.get("consecutive_skips", 0) + 1
skip_stats["explicit_streak"] = True
skip_stats["needed_learns"] = 0
if add_noise_ratio > 0.0 and float(sigma_next) > 0.0 and sigma_up is not None and float(sigma_up) > 0.0:
noise = torch.randn_like(x)
if add_noise_type == "whitened":
noise = (noise - noise.mean()) / (noise.std() + 1e-12)
if official_comfy or alpha_ratio is None or alpha_ratio is True:
x = x + noise * sigma_up
else:
x = alpha_ratio * x + noise * sigma_up
if debug:
try:
x_rms = float(torch.sqrt(torch.mean(x**2)).item())
except Exception:
x_rms = None
print_step_diag(
sampler="gradient_estimation",
step_index=step_index,
sigma_current=sigma_current,
sigma_next=sigma_next,
target_sigma=target_sigma,
sigma_up=sigma_up,
alpha_ratio=alpha_ratio,
h=dt,
c2=None,
b1=None,
b2=None,
eps_norm=hat_norm,
eps_prev_norm=float(torch.norm(prev_eps).item()) if prev_eps is not None else None,
x_rms=x_rms,
flags=f"SKIPPED-{tag}",
)
return x, learning_ratio
else:
if debug:
print(f"gradient_est step {step_index}: explicit skip cancelled (ε̂ invalid: {reason}) hat_norm={hat_norm:.2e}")
else:
if debug:
reason = "need_two_learns_before_skip" if not (es or nl <= 0) else "insufficient_history"
print(f"gradient_est step {step_index}: explicit skip gated ({reason})")
# Decide skip (non-explicit)
if skip_mode == "adaptive":
should_skip, epsilon_hat, meta = decide_skip_adaptive(
epsilon_history=epsilon_history,
step_index=step_index,
total_steps=total_steps,
protect_last_steps=protect_last_steps,
protect_first_steps=protect_first_steps,
skip_stats=skip_stats,
x_current=x,
sigma_current=sigma_current,
sigma_next=target_sigma,
sampler_kind="euler",
anchor_interval=anchor_interval,
max_consecutive_skips=max_consecutive_skips,
)
skip_method = "adaptive"
else:
should_skip, skip_method = should_skip_model_call(1.0, step_index, total_steps, skip_mode, epsilon_history, protect_last_steps, protect_first_steps)
epsilon_hat = None
if should_skip and skip_method is not None:
if epsilon_hat is None:
if skip_method == "richardson":
epsilon_hat = extrapolate_epsilon_richardson(epsilon_history)
elif skip_method == "h4":
epsilon_hat = extrapolate_epsilon_h4(epsilon_history)
else:
epsilon_hat = extrapolate_epsilon_linear(epsilon_history)
prev_eps = epsilon_history[-1] if len(epsilon_history) >= 1 else None
ok, reason, hat_norm, prev_norm = validate_epsilon_hat(epsilon_hat, prev_eps)
if not ok:
if debug:
print(f"gradient_est step {step_index}: skip cancelled (ε̂ invalid: {reason}) hat_norm={hat_norm:.2e}")
else:
if len(epsilon_history) >= 3:
epsilon_hat = epsilon_hat / max(learning_ratio, 1e-8)
d_hat = -(epsilon_hat) / sigma_current
dbar_hat = (ge_gamma - 1.0) * (d_hat - d_prev) if d_prev is not None else 0.0
# Clamp correction magnitude relative to base slope
if isinstance(dbar_hat, torch.Tensor):
try:
_ratio = float(torch.norm(dbar_hat) / (torch.norm(d_hat) + 1e-8))
except Exception:
_ratio = 0.0
if _ratio > 0.25:
dbar_hat = dbar_hat * (0.25 / _ratio)
x = x + (d_hat + (dbar_hat if isinstance(dbar_hat, torch.Tensor) else 0.0)) * dt
if skip_stats is not None:
skip_stats["skipped"] = skip_stats.get("skipped", 0) + 1
skip_stats["consecutive_skips"] = skip_stats.get("consecutive_skips", 0) + 1
skip_stats["explicit_streak"] = True
skip_stats["needed_learns"] = 0
if add_noise_ratio > 0.0 and float(sigma_next) > 0.0 and sigma_up is not None and float(sigma_up) > 0.0:
noise = torch.randn_like(x)
if add_noise_type == "whitened":
noise = (noise - noise.mean()) / (noise.std() + 1e-12)
if official_comfy or alpha_ratio is None or alpha_ratio is True:
x = x + noise * sigma_up
else:
x = alpha_ratio * x + noise * sigma_up
# Ensure x_rms is defined even if debug is False
x_rms = None
if debug:
# Summary line consistent with Euler
print(f"gradient_est step {step_index} [SKIPPED-{skip_method}]: e_norm={hat_norm:.2f}, L={learning_ratio:.4f}, dt={(dt.item() if hasattr(dt, 'item') else float(dt)):.4f}")
try:
x_rms = float(torch.sqrt(torch.mean(x**2)).item())
except Exception:
x_rms = None
print_step_diag(
sampler="gradient_estimation",
step_index=step_index,
sigma_current=sigma_current,
sigma_next=sigma_next,
target_sigma=target_sigma,
sigma_up=sigma_up,
alpha_ratio=alpha_ratio,
h=dt,
c2=None,
b1=None,
b2=None,
eps_norm=hat_norm,
eps_prev_norm=float(torch.norm(prev_eps).item()) if prev_eps is not None else None,
x_rms=x_rms,
flags=f"SKIPPED-{skip_method}",
)
return x, learning_ratio
# REAL Gradient Estimation step
den = model(x, sigma_current * s_in, **extra_args)
d = (x - den) / (sigma_current + 1e-8)
x = x + d * dt
if d_prev is not None:
dbar = (ge_gamma - 1.0) * (d - d_prev)
# Clamp REAL correction for stability
try:
_ratio_real = float(torch.norm(dbar) / (torch.norm(d) + 1e-8))
except Exception:
_ratio_real = 0.0
if _ratio_real > 0.25:
dbar = dbar * (0.25 / _ratio_real)
x = x + dbar * dt
if add_noise_ratio > 0.0 and float(sigma_next) > 0.0 and sigma_up is not None and float(sigma_up) > 0.0:
noise = torch.randn_like(x)
if add_noise_type == "whitened":
noise = (noise - noise.mean()) / (noise.std() + 1e-12)
if official_comfy or alpha_ratio is None or alpha_ratio is True:
x = x + noise * sigma_up
else:
x = alpha_ratio * x + noise * sigma_up
if skip_stats is not None:
skip_stats["model_calls"] = skip_stats.get("model_calls", 0) + 1
skip_stats["consecutive_skips"] = 0
skip_stats["last_anchor_step"] = step_index
# Gating update: REAL call increments learns and may end explicit streak
try:
es = skip_stats.get("explicit_streak", False)
nl = skip_stats.get("needed_learns", 2)
if es:
skip_stats["explicit_streak"] = False
skip_stats["needed_learns"] = 1
else:
skip_stats["needed_learns"] = max(0, int(nl) - 1)
except Exception:
pass
eps_real = den - noisy_latent
epsilon_history.append(eps_real)
if len(epsilon_history) >= 3:
if predictor_type == "h4":
epsilon_hat = extrapolate_epsilon_h4(epsilon_history)
elif predictor_type == "richardson":
epsilon_hat = extrapolate_epsilon_richardson(epsilon_history)
else:
epsilon_hat = extrapolate_epsilon_linear(epsilon_history)
if epsilon_hat is not None:
learn_obs = (torch.norm(epsilon_hat) / (torch.norm(eps_real) + 1e-8)).item()
learning_ratio = smoothing_beta * learning_ratio + (1.0 - smoothing_beta) * learn_obs
learning_ratio = max(0.5, min(2.0, learning_ratio))
if debug:
print(f"gradient_est step {step_index} [LEARN]: learn_obs={learn_obs:.4f}, L={learning_ratio:.4f}, beta={smoothing_beta}")
if debug:
# Summary line consistent with Euler
try:
e_norm = float(torch.norm(eps_real).item())
d_norm = float(torch.norm(d).item())
dt_val = (dt.item() if hasattr(dt, 'item') else float(dt))
except Exception:
e_norm = float('nan'); d_norm = float('nan'); dt_val = float('nan')
print(f"gradient_estimation step {step_index}: e_norm={e_norm:.2f}, d_norm={d_norm:.2f}, dt={dt_val:.4f}, L={learning_ratio:.4f}, beta={smoothing_beta}")
try:
x_rms = float(torch.sqrt(torch.mean(x**2)).item())
except Exception:
x_rms = None
print_step_diag(
sampler="gradient_estimation",
step_index=step_index,
sigma_current=sigma_current,
sigma_next=sigma_next,
target_sigma=target_sigma,
sigma_up=sigma_up,
alpha_ratio=alpha_ratio,
h=dt,
c2=None,
b1=None,
b2=None,
eps_norm=float(torch.norm(eps_real).item()),
eps_prev_norm=float(torch.norm(epsilon_history[-2]).item()) if len(epsilon_history) >= 2 else None,
x_rms=x_rms,
flags="",
)
# Optional SKIP diagnostics for grad-est
try:
if debug and 'd_hat' in locals():
d_norm = float(torch.norm(d_hat).item())
dbar_norm = float(torch.norm(dbar_hat).item()) if isinstance(dbar_hat, torch.Tensor) else 0.0
ratio = dbar_norm / (d_norm + 1e-8)
print(f"gradient_est step {step_index} [SKIP-APPLY]: d_norm={d_norm:.2f}, dbar_norm={dbar_norm:.2f}, ratio={ratio:.2f}, L={learning_ratio:.4f}, gamma={ge_gamma:.2f}")
except Exception:
pass
return x, learning_ratio
|