[yiyixuxu]

Fixes #14388, see the issue for more details on the bug.

This PR:

1. When an InputParam is both required=True and carries a non-None default: InputParam.__post_init__ now keeps the default, treats the input as optional and logs a warning. This matches what the code has always done at runtime, so no pipeline changes behavior. The usual way to hit this is InputParam.template("num_inference_steps", required=True): when the user try to reuse the template for its type hint and description, but want to mark it as required.

2. Go through every input in the library that was declared both ways and decide explicitly which one it means:

   • num_inference_steps:
     • the set-timesteps blocks keep the plain template default, i.e. optional with default 50: Cosmos3SetTimestepsStep, Cosmos3TransferSetTimestepsStep, MiniMaxH3SetTimestepsStep
     • every block that runs after set-timesteps declares required=True, default=None, since by then the value is always in the state: Cosmos3DenoiseLoopWrapper, Krea2LoopDenoiser, Krea2DenoiseLoopWrapper, LTXLoopDenoiser, LTXImage2VideoLoopDenoiser, LTXDenoiseLoopWrapper, LTX2LoopDenoiser, LTX2DenoiseLoopWrapper, QwenImageDenoiseLoopWrapper, QwenImage21DenoiseLoopWrapper
   • dtype (LTX, LTX-2) and batch_size (LTX): required=True, default=None. They are intermediates always produced upstream, so the default was never used.

Also re-generated the auto docstrings and added a gotcha for the modular tests: I noticed the MiniMax-H3 workflow test explicitly listed num_inference_steps as a required input. That test is meant to be a snapshot of the pipeline's contract (components, configs, inputs with their defaults) so deviations like this get caught, which only works if the table is read back against the other pipelines rather than copied from the code.

[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.

[Chessing234]

the pitfalls list in .ai/references/modular.md numbers two items as 10. the torch.randn note never moved to 11, so the helpers note is 11 instead of 12.