[docs] Fix missing parameters in docstrings

[stevhliu]

Fixes #10416 by adding missing parameters in FlowMatchEulerDiscreteScheduler and DPMSolverMultistepScheduler .

[stevhliu]

Can I get some help filling in what this parameter does? @hlky

[hlky]

use_flow_sigmas was added with Sana, it is to support Flow Match in the existing schedulers, flow_shift is like shift in FlowMatchEuler. Schedulers with use_flow_sigmas have limited support for other models as some don't support passing sigmas or dynamic shifting, as I understand the schedulers are in use by some integrations with Flux and SD3 but there may be quality issues with outputs.

diffusers/src/diffusers/schedulers/scheduling_unipc_multistep.py

Line 301 in 91008aa

def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):

diffusers/src/diffusers/schedulers/scheduling_unipc_multistep.py

Lines 380 to 382 in 91008aa

	alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1)
	sigmas = 1.0 - alphas
	sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1].copy()

It's similar in FlowMatchEuler, by default it is for SD3 and other models pass sigmas

diffusers/src/diffusers/schedulers/scheduling_flow_match_euler_discrete.py

Lines 238 to 240 in 91008aa

	timesteps = np.linspace(
	self._sigma_to_t(self.sigma_max), self._sigma_to_t(self.sigma_min), num_inference_steps
	)

diffusers/src/diffusers/pipelines/hunyuan_video/pipeline_hunyuan_video.py

Line 608 in 91008aa

sigmas = np.linspace(1.0, 0.0, num_inference_steps + 1)[:-1] if sigmas is None else sigmas

diffusers/src/diffusers/pipelines/mochi/pipeline_mochi.py

Line 657 in 91008aa

sigmas = linear_quadratic_schedule(num_inference_steps, threshold_noise)

diffusers/src/diffusers/pipelines/flux/pipeline_flux.py

Line 823 in 91008aa

sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas

The schedulers with use_flow_sigmas are scheduling_deis_multistep, scheduling_dpmsolver_multistep_inverse, scheduling_dpmsolver_multistep, scheduling_dpmsolver_singlestep, scheduling_sasolver, scheduling_unipc_multistep.

[stevhliu]

Also need some help with this parameter please :)

[hlky]

invert_sigmas was added with Mochi, the model works in the opposite way than usual so the option flips sigmas after calculation

diffusers/src/diffusers/schedulers/scheduling_flow_match_euler_discrete.py

Lines 268 to 273 in 91008aa

	if self.config.invert_sigmas:
	sigmas = 1.0 - sigmas
	timesteps = sigmas * self.config.num_train_timesteps
	sigmas = torch.cat([sigmas, torch.ones(1, device=sigmas.device)])
	else:
	sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)])

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[stevhliu]

Sorry I wasn't clear enough, but I was referring to the shift_terminal parameter. Let me know what you think 🙂

[hlky]

Stretches and shifts the timestep schedule to ensure it terminates at the shift_terminal value. Used by LTXVideo.

[stevhliu]

Thanks! Let me know if the descriptions are clear enough, otherwise I think we can merge