feat: add pruner algroithm #470
Conversation
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
|
|
||
| prompts = self._extract_prompts(args, kwargs) | ||
| max_num_tokens = max(len(self.tokenizer.encode(p)) for p in prompts) | ||
|
|
There was a problem hiding this comment.
Bug: Fails when prompts are not strings
wrapped_call computes max_num_tokens via max(... for p in prompts), but _extract_prompts can return an empty list (e.g., callers pass prompt_embeds/no prompt strings). This triggers a runtime ValueError and breaks inference for valid diffusers call patterns.
| sequence_length *= 2 | ||
| if sequence_length >= max_sequence_length: | ||
| sequence_length = max_sequence_length | ||
| kwargs["max_sequence_length"] = sequence_length |
There was a problem hiding this comment.
Bug: Assumes max_sequence_length has a default
default_max_sequence_length is read from inspect.signature(...).parameters["max_sequence_length"].default. If that default is inspect._empty (or the user passes max_sequence_length=None), comparisons like sequence_length >= max_sequence_length can raise TypeError, causing runtime failures on pipelines with a required max_sequence_length.
| """Disable prompt pruning by unwrapping the pipe.""" | ||
| if self.pipe_call: | ||
| self.pipe.__call__ = self.pipe_call | ||
|
|
There was a problem hiding this comment.
Bug: Helper disable may crash before enable
disable accesses self.pipe_call without initializing it in __init__. If disable is called before enable/wrap_pipe runs (or after partial initialization failures), this can raise AttributeError and prevent cleanup/saving flows that call .disable() on helpers.
minettekaum
changed the title
Description
I added
padding_pruningalgorithm andflux_tiny_random_with_tokenizerto fixtures to be able to testpadded_pruningpadding_pruningalgorithm insrc/pruna/algorithms/padding_pruning.pypadding_pruningalgorithm intests/algorithms/testers/padding_pruning.pyflux_tiny_random_with_tokenizeradded intests/fixtures.pyso thatpadding_pruningalgorithm can be tested.Type of Change
How Has This Been Tested?
Yes, by tests for
padding_pruning.Checklist