Modular pipelines: warn when a denoiser_input_fields entry is not passed to the transformer #15022

The loop denoisers forward the fields tagged denoiser_input_fields whose name matches the transformer's forward signature and drop the rest silently, e.g. in qwenimage/denoise.py:

transformer_args = set(inspect.signature(components.transformer.forward).parameters.keys())
for field_name, field_value in block_state.denoiser_input_fields.items():
    if field_name in transformer_args and field_name not in guider_inputs:
        additional_cond_kwargs[field_name] = field_value

So a custom block tagging a field under a name the transformer does not take (a typo, or a field meant for another transformer) runs without any error and the conditioning never reaches the model. Same pattern in the helios, wan, wan_animate_2, ltx2 and minimax_h3 denoisers.

Proposal: work on a copy of the dict, pop the fields the block consumes by name and the ones matching the signature, and logger.warning_once whatever is left:

`denoiser_input_fields` not passed to the transformer: ['typo_field']

Two things found while trying this out:

• QwenImage's edit-plus and layered RoPE input steps still tag txt_seq_lens / negative_txt_seq_lens for the denoiser, but the transformer no longer takes them since [core] remove txt_seq_lens from qwen transformer. #13674.
• diffusers' logger has no warning_once of its own. The existing logger.warning_once calls only work because transformers patches logging.Logger when it is imported.

Posted by Claude Code on behalf of @yiyixuxu, who reviewed and approved this text.