Repository navigation
Implement missing features on ModularPipeline #12386
Description
Activity
This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.
Please note that issues that do not follow the contributing guidelines are likely to be ignored.
we have input_names property now
https://github.com/huggingface/diffusers/blob/main/src/diffusers/modular_pipelines/modular_pipeline.py#L504also there is required_inputs https://github.com/huggingface/diffusers/blob/main/src/diffusers/modular_pipelines/modular_pipeline.py#L285
we will add support for callback etc
This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.
Please note that issues that do not follow the contributing guidelines are likely to be ignored.
diffusers is moving to modular-only for new pipelines, but modular pipelines still dont have clear way to perform callbacks more than 6 months later?
cc @sayakpaul- added a commit that references this issue
on Oct 6, 2026 hi hi
we will include this PR in next release #14159
it adds first-class streaming to modulaar pipelines, so you won't need the callback/interrupt mechanism from the standard pipeline: pipe.stream() let you work with the pipeline interactively:for event in pipe.stream(prompt=...): latent = event.state.get("latents")
as i'm looking to take advantage of new
ModularPipelineask is to implement some currently missing featuresmy use case is to convert existing loaded model using standard pipeline into modular pipeline. that functionality was provided via #11915 and is now working.
first minor obstacle is that modular pipeline does not have defined params for execution
in standard pipeline i can inspect
__call__signature to see which are allowed paramsi currently work around this using
possible = [input_param.name for input_param in model.blocks.inputs]please advise if this is acceptable
second one is that modular pipelines don't seem to implement normal callbacks at all (e.g.
callback_on_step_end_tensor_inputs? at the minimum we need some kind of callback functionality to capture interim latents on each stepthird is more cosmetic - modular pipeline does implement
set_progress_bar_config, but its not doing anything as its not implement on actual block (tested withStableDiffusionXLModularPipeline)cc @yiyixuxu @DN6 @sayakpaul