Modular pipelines: variant blocksets sharing a `model_name` init the wrong pipeline class

[yiyixuxu]

ModularPipelineBlocks.init_pipeline() resolves the pipeline class through MODULAR_PIPELINE_MAPPING[self.model_name], calling the map fn without a config. The config-branching map fns (_flux2_klein_map_fn, _krea2_map_fn, _helios_pyramid_map_fn, _wan_map_fn, _wan_i2v_map_fn) can only pick the variant class when given a repo config (the model_index.json fallback in from_pretrained), so from the blocks side they always return their None default. Any variant blockset that shares its base's model_name inits as the wrong pipeline class:

variant blockset	model_name	init_pipeline() returns	should be
Flux2KleinBaseAutoBlocks	flux2-klein	Flux2KleinModularPipeline (the distilled default)	Flux2KleinBaseModularPipeline
Krea2TurboAutoBlocks	krea2	Krea2ModularPipeline	Krea2TurboModularPipeline
HeliosPyramidDistilledAutoBlocks	helios-pyramid	HeliosPyramidModularPipeline	HeliosPyramidDistilledModularPipeline
Wan22Blocks	wan	WanModularPipeline	Wan22ModularPipeline
Wan22Image2VideoBlocks	wan-i2v	WanImage2VideoModularPipeline	Wan22Image2VideoModularPipeline
Cosmos3DistilledBlocks	cosmos3-omni	Cosmos3OmniModularPipeline	Cosmos3DistilledModularPipeline
LTX25AutoBlocks	ltx2	LTX2ModularPipeline	LTX25ModularPipeline

Consequences:

• Krea2 turbo is functionally wrong: Krea2TurboModularPipeline pins requires_unconditional_embeds = False; the base class consults the guider, so a turbo pipeline built from blocks can be asked to encode negative embeds.
• The error is sticky: save_pretrained writes _class_name from the live pipeline class, so a wrongly-inited variant round-trips the wrong class through its saved repo.
• Most variants only override default_blocks_name, so the damage is usually latent — but the fast test suites for krea2-turbo, flux2-klein-base, and cosmos3-distilled all go through init_pipeline() and currently exercise the base class.
• Cosmos3DistilledBlocks has no config branch at all, so even the model_index.json path can never return the distilled class; the "ltx2.5" mapping key is dead code (no blockset declares that name).

Proposed fix — the pattern qwenimage (4 names) and wan-animate-2 (wan-animate-2 / wan-animate-2-distilled, #14413) already follow: give each variant blockset its own model_name with its own _create_default_map_fn entry. It's a small change now that every variant has its own blockset file, and it also fixes the save/reload round-trip. The config-branching map fns stay for the model_index.json fallback in from_pretrained (that path derives one model name per standard pipeline class, so it still needs the config to tell variants apart).

The same PR should add a guard test to ModularPipelineTesterMixin so this can't regress:

def test_init_pipeline_class(self):
    pipe = self.pipeline_blocks_class().init_pipeline()
    assert type(pipe) is self.pipeline_class

Every tester already declares pipeline_class, so this is immediately red today for the krea2-turbo, flux2-klein-base, and cosmos3-distilled suites and goes green with the fix.

---

Posted by Claude (Claude Code) on behalf of @yiyixuxu, as a reminder from the Wan-Animate-2 integration work (#14413).