Stability-AI/generative-models · error · ValueError

Order {order} too high for step {i}

Error message

Order {order} too high for step {i}

What it means

linear_multistep_coeff builds the linear multistep coefficient for order-order solvers; it needs `order-1` previous steps, so calling it at step i with order-1 > i raises ValueError.

Source

Thrown at sgm/modules/diffusionmodules/sampling_utils.py:9

import torch
from scipy import integrate

from ...util import append_dims


def linear_multistep_coeff(order, t, i, j, epsrel=1e-4):
    if order - 1 > i:
        raise ValueError(f"Order {order} too high for step {i}")

    def fn(tau):
        prod = 1.0
        for k in range(order):
            if j == k:
                continue
            prod *= (tau - t[i - k]) / (t[i - j] - t[i - k])
        return prod

    return integrate.quad(fn, t[i], t[i + 1], epsrel=epsrel)[0]


def get_ancestral_step(sigma_from, sigma_to, eta=1.0):
    if not eta:
        return sigma_to, 0.0
    sigma_up = torch.minimum(
        sigma_to,
        eta

View on GitHub (pinned to e8cd657656)

Solutions

1. Use min(order, i+1) as the effective order for the first steps in the sampling loop
2. Lower the sampler order (e.g. order=2 or 3) in the sampling config
3. Ensure the sampler's ramp-up logic (self.rampup / get_order) is invoked rather than calling linear_multistep_coeff directly with a fixed order

Example fix

// before
coef = linear_multistep_coeff(4, t, i, old_order)
// after
order = min(4, i + 1)
coef = linear_multistep_coeff(order, t, i, old_order)

def assert_order_fits(order, i):
    if order - 1 > i:
        raise ValueError(f"order {order} needs {order-1} prior steps, but i={i}")
assert_order_fits(sampler.order, step_index)

def order_fits(order: int, i: int) -> bool:
    return order - 1 <= i

try:
    x = sampler.sample(S, shape, order=4)
except ValueError as e:
    if "too high for step" in str(e):
        x = sampler.sample(S, shape, order=2)
    else:
        raise

When it happens

Trigger: Running a linear multistep sampler (e.g. LDMS) with order > i+1 at early steps — e.g. sampler order=4 while ramping starts at i=0 and the sampler does not reduce order for the first steps.

Common situations: Custom sampling loops calling LinearMultistepCoeff/Sampler.sample with a high order and too few warm-up steps; misconfigured sampler args (order larger than allowed for the step schedule).

Related errors

• unknown merge strategy {self.merge_strategy}
• Decay must be between 0 and 1
• unknown merge strategy {merge_strategy}
• rearranging not available for {len(in_shape)}-dimensional in
• Unknown loss type {self.loss_type}

AI-assisted analysis of Stability-AI/generative-models@e8cd657656 (2026-08-29). Data as JSON: /api/errors/b60146b39ab925f2. Report an issue: GitHub.