Issues with FlowMatchEulerDiscreteScheduler.set_timesteps()

[dxqbYD]

Describe the bug

Why does num_inference_steps have the default None? It's not an Optional. It cannot be None. This leads to weird error messages if you skip this parameter.

diffusers/src/diffusers/schedulers/scheduling_flow_match_euler_discrete.py

Line 239 in 37c9697

num_inference_steps: int = None,

sigmas is undocumented:

diffusers/src/diffusers/schedulers/scheduling_flow_match_euler_discrete.py

Line 241 in 37c9697

sigmas: Optional[List[float]] = None,

mu is undocumented, even though it can be a required parameter (depending on configuration):

diffusers/src/diffusers/schedulers/scheduling_flow_match_euler_discrete.py

Line 242 in 37c9697

mu: Optional[float] = None,

Reproduction

see above

Logs

System Info

HEAD

Who can help?

@yiyixuxu

[hlky]

Why does num_inference_steps have the default None? It's not an Optional. It cannot be None

num_inference_steps is optional, it is None when we pass sigmas.

This leads to weird error messages if you skip this parameter.

What is the error message?

sigmas is undocumented

It appears to be missing from the docstring, sigmas is passed to FlowMatchEulerDiscreteScheduler to accommodate different "base schedules" which tend to be unique in FlowMatch models. The default in FlowMatchEulerDiscreteScheduler, as in, when we pass num_inference_steps is for SD3 as this was added first. Flux, HunyuanVideo, LTX, Mochi all pass their own base schedule from through sigmas to FlowMatchEulerDiscreteScheduler from the pipeline.

mu is undocumented, even though it can be a required parameter

mu is used when use_dynamic_shifting=True, there is a check that mu is not None when use_dynamic_shifting=True.

[dxqbYD]

error message:

 scheduler.set_timesteps()
  File "...\venv\src\diffusers\src\diffusers\schedulers\scheduling_flow_match_euler_discrete.py", line 257, in set_timesteps
    timesteps = np.linspace(
  File "...\venv\lib\site-packages\numpy\core\function_base.py", line 122, in linspace
    num = operator.index(num)
TypeError: 'NoneType' object cannot be interpreted as an integer

Thank you for your explanations regarding the other parameters, but I meant that they should be documented in the code/docs

[hlky]

I see. One of the arguments must be used.