Scheduler `step()` methods silently ignore documented sampling parameters (`s_churn`, `s_noise`, `eta`, `generator`, `variance_noise`)

[4ktLuffy]

Describe the bug

Four scheduler step() methods accept parameters that are documented in their Args: block but never referenced in the function body. Passing them has no effect and no warning is raised. Filing as one systematic issue per the fix patterns, not one-offs guideline.

File	Method	Accepted but never used
scheduling_flow_match_euler_discrete.py:423	step()	s_churn, s_tmin, s_tmax, s_noise
scheduling_ddim_cogvideox.py:326	step()	eta, use_clipped_model_output, generator, variance_noise
scheduling_dpm_cogvideox.py:401	step()	eta, use_clipped_model_output, variance_noise
scheduling_helios.py:311	step_euler()	generator

Why this is user-visible

• generator is discarded in DDIMSchedulerCogVideoX.step() and HeliosScheduler.step_euler(). A caller passing a seeded torch.Generator for reproducibility gets no error and no reproducibility.
• eta is discarded in both CogVideoX schedulers. eta is the DDIM stochasticity control (0 = deterministic, 1 = DDPM); setting it does nothing.
• s_noise is discarded in FlowMatchEulerDiscreteScheduler.step(), where it is documented as "Scaling factor for noise added to the sample". The stochastic_sampling branch calls randn_tensor(...) and never applies it:

if self.config.stochastic_sampling:
    x0 = sample - current_sigma * model_output
    noise = randn_tensor(sample.shape, generator=generator, device=sample.device, dtype=sample.dtype)
    prev_sample = (1.0 - next_sigma) * x0 + next_sigma * noise   # s_noise never applied

s_churn, s_tmin and s_tmax in that same signature also have empty docstring descriptions, which suggests they were copied from EulerDiscreteScheduler without being wired up.

How this was found

An AST scan comparing each function's parameter list against every ast.Name referenced in its body, with the signature and docstring excluded so documentation mentions don't count as usage. Verified per-function rather than by grep. Happy to share the script.

Which fix do you want?

Two defensible directions, and I'd rather not guess:

1. Remove the dead parameters — matches the AGENTS.md guidance ("do not carry unused method parameters 'for API consistency'"), but changes a public signature.
2. Implement them — s_noise in particular reads as a missing implementation rather than dead weight, since the stochastic branch it belongs to does exist.

A third option would be raising on non-default values instead of ignoring them, per "raise a concise error for unsupported cases".

Happy to open a PR once a maintainer confirms both the scope and which direction you'd prefer.

[4ktLuffy]

Correction to the table above: the class is CogVideoXDDIMScheduler (and CogVideoXDPMScheduler), not DDIMSchedulerCogVideoX. File paths and line numbers are unaffected.

Runnable reproduction on main — CPU only, no model weights:

import torch
from diffusers import CogVideoXDDIMScheduler, FlowMatchEulerDiscreteScheduler

torch.manual_seed(0)
model_output, sample = torch.randn(1, 4, 8, 8), torch.randn(1, 4, 8, 8)

# s_noise: same generator seed, three different s_noise values
outs = []
for s_noise in (1.0, 0.0, 1000.0):
    s = FlowMatchEulerDiscreteScheduler(stochastic_sampling=True)
    s.set_timesteps(4)
    g = torch.Generator().manual_seed(42)
    outs.append(s.step(model_output, s.timesteps[0], sample, s_noise=s_noise, generator=g).prev_sample)
print("s_noise ignored:", torch.equal(outs[0], outs[1]) and torch.equal(outs[0], outs[2]))

# generator / eta
def run(eta, seed):
    s = CogVideoXDDIMScheduler(); s.set_timesteps(4)
    g = torch.Generator().manual_seed(seed)
    return s.step(model_output, s.timesteps[0], sample, eta=eta, generator=g).prev_sample
print("generator ignored:", torch.equal(run(0.0, 42), run(0.0, 12345)))
print("eta ignored:      ", torch.equal(run(0.0, 42), run(1.0, 42)))

Output on 0.40.0.dev0, torch 2.13.0:

s_noise ignored: True
generator ignored: True
eta ignored:       True

s_churn=0 vs s_churn=100 is likewise bit-identical.

[4ktLuffy]

Some additional evidence on the "remove vs implement" question, since it turned out to be more than tidiness.

The declarations cause a false positive in capability detection. Pipelines decide whether to forward eta and generator by inspecting the scheduler's signature — prepare_extra_step_kwargs in pipeline_stable_diffusion.py, copied into pipeline_cogvideox.py among others:

accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
if accepts_eta:
    extra_step_kwargs["eta"] = eta

accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
if accepts_generator:
    extra_step_kwargs["generator"] = generator

CogVideoXPipeline is typed for CogVideoXDDIMScheduler | CogVideoXDPMScheduler, both of which declare eta and generator. So the pipeline detects them, forwards them on every step, and the scheduler discards them:

import inspect, torch
from diffusers import CogVideoXDDIMScheduler

params = set(inspect.signature(CogVideoXDDIMScheduler.step).parameters)
print("accepts_eta      ", "eta" in params)         # True
print("accepts_generator", "generator" in params)   # True

torch.manual_seed(0)
mo, s = torch.randn(1, 4, 8, 8), torch.randn(1, 4, 8, 8)
def run(eta, seed):
    sch = CogVideoXDDIMScheduler(); sch.set_timesteps(4)
    return sch.step(mo, sch.timesteps[0], s, eta=eta,
                    generator=torch.Generator().manual_seed(seed)).prev_sample
print("identical across generators:", torch.equal(run(0.0, 1), run(0.0, 999)))  # True
print("identical across eta:       ", torch.equal(run(0.0, 1), run(1.0, 1)))    # True

The practical effect is that a user passing a seeded generator through CogVideoXPipeline gets no error, no warning, and no reproducibility.

The flow-match case looks like a copy that lost its body. EulerDiscreteScheduler.step — the signature FlowMatchEulerDiscreteScheduler.step mirrors — implements all four:

gamma = min(s_churn / (len(self.sigmas) - 1), 2**0.5 - 1) if s_tmin <= sigma <= s_tmax else 0.0
eps = noise * s_noise

FlowMatchEulerDiscreteScheduler has the signature and the docstring but none of that logic, and s_churn / s_tmin / s_tmax have empty docstring descriptions there. (This one isn't reached through prepare_extra_step_kwargs, which only forwards eta and generator — it affects direct callers.)

Why this argues for removal. Dropping the parameters makes accepts_eta / accepts_generator correctly return False, so pipelines stop forwarding them. Runtime behavior is unchanged — they were already ignored — but introspection stops reporting a capability that isn't there. Implementing them instead means adding real sampling math to four schedulers, which needs your intent rather than my guess.

Happy to open a PR for the removal once you confirm the direction, or to leave it if you'd rather keep the signatures for compatibility.